diff --git "a/openvino_model.xml" "b/openvino_model.xml" new file mode 100644--- /dev/null +++ "b/openvino_model.xml" @@ -0,0 +1,31342 @@ +<?xml version="1.0"?> +<net name="torch_jit" version="11"> + <layers> + <layer id="1" name="input_ids" type="Parameter" version="opset1"> + <data shape="?,?" element_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="input_ids" /> + </rt_info> + <output> + <port id="0" precision="I64" names="input_ids"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="0" name="attention_mask" type="Parameter" version="opset1"> + <data shape="?,?" element_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="attention_mask" /> + </rt_info> + <output> + <port id="0" precision="I64" names="attention_mask"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="2" name="roberta.encoder.layer.11.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="0" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="3" name="roberta.encoder.layer.11.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="3072" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="4" name="roberta.encoder.layer.11.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="15360" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="5" name="roberta.encoder.layer.11.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="18432" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="6" name="roberta.encoder.layer.10.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="21504" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="7" name="roberta.encoder.layer.10.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="24576" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="8" name="roberta.encoder.layer.10.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="36864" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="9" name="roberta.encoder.layer.10.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="39936" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="10" name="roberta.encoder.layer.9.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="43008" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="11" name="roberta.encoder.layer.9.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="46080" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="12" name="roberta.encoder.layer.9.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="58368" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="13" name="roberta.encoder.layer.9.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="61440" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="14" name="roberta.encoder.layer.8.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="64512" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="15" name="roberta.encoder.layer.8.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="67584" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="16" name="roberta.encoder.layer.8.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="79872" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="17" name="roberta.encoder.layer.8.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="82944" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="18" name="roberta.encoder.layer.7.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="86016" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="19" name="roberta.encoder.layer.7.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="89088" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="20" name="roberta.encoder.layer.7.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="101376" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="21" name="roberta.encoder.layer.7.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="104448" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="22" name="roberta.encoder.layer.6.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="107520" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="23" name="roberta.encoder.layer.6.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="110592" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="24" name="roberta.encoder.layer.6.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="122880" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="25" name="roberta.encoder.layer.6.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="125952" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="26" name="roberta.encoder.layer.5.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="129024" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="27" name="roberta.encoder.layer.5.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="132096" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="28" name="roberta.encoder.layer.5.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="144384" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="29" name="roberta.encoder.layer.5.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="147456" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="30" name="roberta.encoder.layer.4.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="150528" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="31" name="roberta.encoder.layer.4.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="153600" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="32" name="roberta.encoder.layer.4.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="165888" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="33" name="roberta.encoder.layer.4.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="168960" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="34" name="roberta.encoder.layer.3.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="172032" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="35" name="roberta.encoder.layer.3.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="175104" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="36" name="roberta.encoder.layer.3.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="187392" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="37" name="roberta.encoder.layer.3.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="190464" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="38" name="roberta.encoder.layer.2.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="193536" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="39" name="roberta.encoder.layer.2.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="196608" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="40" name="roberta.encoder.layer.2.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="208896" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="41" name="roberta.encoder.layer.2.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="211968" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="42" name="roberta.encoder.layer.1.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="215040" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="43" name="roberta.encoder.layer.1.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="218112" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="44" name="roberta.encoder.layer.1.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="230400" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="45" name="roberta.encoder.layer.1.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="233472" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="46" name="roberta.encoder.layer.0.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="236544" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="47" name="roberta.encoder.layer.0.intermediate.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="3072" offset="239616" size="12288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.intermediate.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.intermediate.dense.bias"> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="48" name="roberta.encoder.layer.0.attention.output.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="251904" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.attention.output.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.attention.output.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="49" name="roberta.encoder.layer.0.attention.self.query.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="254976" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.attention.self.query.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.attention.self.query.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="50" name="roberta.embeddings.word_embeddings.weight" type="Const" version="opset1"> + <data element_type="f32" shape="901629, 768" offset="258048" size="2769804288" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.embeddings.word_embeddings.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.embeddings.word_embeddings.weight"> + <dim>901629</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="51" name="Constant_281" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_281" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="52" name="/roberta/embeddings/word_embeddings/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/word_embeddings/Gather" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>901629</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="FP32" names="/roberta/embeddings/word_embeddings/Gather_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="53" name="roberta.embeddings.token_type_embeddings.weight" type="Const" version="opset1"> + <data element_type="f32" shape="1, 768" offset="2770062344" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.embeddings.token_type_embeddings.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.embeddings.token_type_embeddings.weight"> + <dim>1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="54" name="Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1, 514" offset="2770065416" size="4112" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="onnx::Slice_212"> + <dim>1</dim> + <dim>514</dim> + </port> + </output> + </layer> + <layer id="55" name="/roberta/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="56" name="/roberta/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/Shape_1_output_0"> + <dim>2</dim> + </port> + </output> + </layer> + <layer id="57" name="/roberta/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/Constant_2_output_0" /> + </output> + </layer> + <layer id="58" name="Constant_211" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_211" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="59" name="/roberta/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_2, /roberta/Gather_1, Constant_211" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>2</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/Gather_1_output_0" /> + </output> + </layer> + <layer id="60" name="Constant_215" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_215" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="61" name="/roberta/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Unsqueeze, Constant_215" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="62" name="/roberta/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/Constant_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="63" name="/roberta/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/Constant_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="64" name="/roberta/Slice" type="Slice" version="opset8"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Slice" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + <dim>514</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + <port id="4" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="5" precision="I64" names="/roberta/Slice_output_0"> + <dim>1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="65" name="/roberta/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Shape" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/Shape_output_0"> + <dim>2</dim> + </port> + </output> + </layer> + <layer id="66" name="/roberta/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/Constant_1_output_0" /> + </output> + </layer> + <layer id="67" name="Constant_207" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_207" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="68" name="/roberta/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_1, /roberta/Gather, Constant_207" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>2</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/Gather_output_0" /> + </output> + </layer> + <layer id="69" name="Constant_226" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_226" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="70" name="/roberta/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Unsqueeze_1, Constant_226" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="71" name="Constant_228" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_228" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="72" name="/roberta/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Unsqueeze_2, Constant_228" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="73" name="/roberta/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Concat" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Concat_output_0"> + <dim>2</dim> + </port> + </output> + </layer> + <layer id="74" name="Constant_5933" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Reshape" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="75" name="/roberta/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Reshape" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>2</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Reshape_output_0"> + <dim>2</dim> + </port> + </output> + </layer> + <layer id="76" name="Constant_237" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_237" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="77" name="/roberta/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>2</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/Shape_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="78" name="/roberta/ConstantOfShape" type="Broadcast" version="opset3"> + <data mode="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/ConstantOfShape, Constant_237" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/ConstantOfShape_output_0"> + <dim>2</dim> + </port> + </output> + </layer> + <layer id="79" name="/roberta/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/Constant_7_output_0" /> + </output> + </layer> + <layer id="80" name="/roberta/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_7, /roberta/Mul" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>2</dim> + </port> + <port id="1" precision="I64" /> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Mul_output_0"> + <dim>2</dim> + </port> + </output> + </layer> + <layer id="81" name="/roberta/Equal" type="Equal" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Equal" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>2</dim> + </port> + <port id="1" precision="I64"> + <dim>2</dim> + </port> + </input> + <output> + <port id="2" precision="BOOL" names="/roberta/Equal_output_0"> + <dim>2</dim> + </port> + </output> + </layer> + <layer id="82" name="/roberta/Where" type="Select" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Where" /> + </rt_info> + <input> + <port id="0" precision="BOOL"> + <dim>2</dim> + </port> + <port id="1" precision="I64"> + <dim>2</dim> + </port> + <port id="2" precision="I64"> + <dim>2</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/Where_output_0"> + <dim>2</dim> + </port> + </output> + </layer> + <layer id="83" name="/roberta/Expand" type="Broadcast" version="opset3"> + <data mode="bidirectional" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Expand" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="I64"> + <dim>2</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Expand_output_0"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="84" name="Constant_283" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_283" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="85" name="/roberta/embeddings/token_type_embeddings/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/token_type_embeddings/Gather" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="FP32" names="/roberta/embeddings/token_type_embeddings/Gather_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="86" name="/roberta/embeddings/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="87" name="roberta.embeddings.position_embeddings.weight" type="Const" version="opset1"> + <data element_type="f32" shape="514, 768" offset="2770069544" size="1579008" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.embeddings.position_embeddings.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.embeddings.position_embeddings.weight"> + <dim>514</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="88" name="/roberta/embeddings/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/embeddings/Constant_output_0" /> + </output> + </layer> + <layer id="89" name="/roberta/embeddings/Equal" type="Equal" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Equal" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="I64" /> + </input> + <output> + <port id="2" precision="BOOL" names="/roberta/embeddings/Equal_output_0"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="90" name="/roberta/embeddings/Not" type="LogicalNot" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Not" /> + </rt_info> + <input> + <port id="0" precision="BOOL"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="BOOL" names="/roberta/embeddings/Not_output_0"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="91" name="/roberta/embeddings/Cast" type="Convert" version="opset1"> + <data destination_type="i32" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Cast" /> + </rt_info> + <input> + <port id="0" precision="BOOL"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="I32" names="/roberta/embeddings/Cast_output_0"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="92" name="/roberta/embeddings/Constant_1" type="Const" version="opset1"> + <data element_type="i32" shape="" offset="2771648552" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I32" names="/roberta/embeddings/Constant_1_output_0" /> + </output> + </layer> + <layer id="93" name="/roberta/embeddings/CumSum" type="CumSum" version="opset3"> + <data exclusive="false" reverse="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/CumSum" /> + </rt_info> + <input> + <port id="0" precision="I32"> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="I32" /> + </input> + <output> + <port id="2" precision="I32" names="/roberta/embeddings/CumSum_output_0"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="94" name="/roberta/embeddings/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Mul" /> + </rt_info> + <input> + <port id="0" precision="I32"> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="I32"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="I32" names="/roberta/embeddings/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="95" name="/roberta/embeddings/Cast_1" type="Convert" version="opset1"> + <data destination_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Cast_1" /> + </rt_info> + <input> + <port id="0" precision="I32"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/embeddings/Cast_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="96" name="/roberta/embeddings/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/embeddings/Constant_2_output_0" /> + </output> + </layer> + <layer id="97" name="/roberta/embeddings/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Add" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="I64" /> + </input> + <output> + <port id="2" precision="I64" names="/roberta/embeddings/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="98" name="Constant_286" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_286" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="99" name="/roberta/embeddings/position_embeddings/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/position_embeddings/Gather" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>514</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="FP32" names="/roberta/embeddings/position_embeddings/Gather_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="100" name="/roberta/embeddings/Add_2" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/Add_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/Add_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="101" name="Constant_289" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_289" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="102" name="/roberta/embeddings/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="103" name="/roberta/embeddings/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="104" name="/roberta/embeddings/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/embeddings/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="105" name="/roberta/embeddings/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="106" name="Constant_294" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_294" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="107" name="/roberta/embeddings/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="108" name="/roberta/embeddings/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/embeddings/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="109" name="/roberta/embeddings/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="110" name="/roberta/embeddings/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/embeddings/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="111" name="/roberta/embeddings/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="112" name="roberta.embeddings.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2771648564" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.embeddings.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.embeddings.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="113" name="/roberta/embeddings/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="114" name="roberta.embeddings.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2771651636" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.embeddings.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.embeddings.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="115" name="/roberta/embeddings/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/embeddings/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/embeddings/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="116" name="onnx::MatMul_1556" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2771654708" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1556" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1556"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="117" name="/roberta/encoder/layer.0/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="118" name="/roberta/encoder/layer.0/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="119" name="/roberta/encoder/layer.0/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.0/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="120" name="/roberta/encoder/layer.0/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="121" name="Constant_364" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_364" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="122" name="/roberta/encoder/layer.0/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_8, /roberta/encoder/layer.0/attention/self/Gather_4, Constant_364" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="123" name="Constant_370" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_370" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="124" name="/roberta/encoder/layer.0/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Unsqueeze_4, Constant_370" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.0/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="125" name="/roberta/encoder/layer.0/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.0/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="126" name="/roberta/encoder/layer.0/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="127" name="Constant_368" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_368" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="128" name="/roberta/encoder/layer.0/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_9, /roberta/encoder/layer.0/attention/self/Gather_5, Constant_368" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="129" name="Constant_372" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_372" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="130" name="/roberta/encoder/layer.0/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Unsqueeze_5, Constant_372" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.0/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="131" name="/roberta/encoder/layer.0/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="132" name="/roberta/encoder/layer.0/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="133" name="/roberta/encoder/layer.0/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Concat_2, /roberta/encoder/layer.0/attention/self/Constant_10, /roberta/encoder/layer.0/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.0/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="134" name="/roberta/encoder/layer.0/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="135" name="Constant_388" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_388" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="136" name="/roberta/encoder/layer.0/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="137" name="roberta.encoder.layer.0.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2774014052" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="138" name="onnx::MatMul_1557" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2774017124" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1557" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1557"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="139" name="/roberta/encoder/layer.0/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="140" name="/roberta/encoder/layer.0/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="141" name="/roberta/encoder/layer.0/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.0/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="142" name="/roberta/encoder/layer.0/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="143" name="Constant_308" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_308" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="144" name="/roberta/encoder/layer.0/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant, /roberta/encoder/layer.0/attention/self/Gather, Constant_308" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="145" name="Constant_314" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_314" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="146" name="/roberta/encoder/layer.0/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Unsqueeze, Constant_314" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.0/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="147" name="/roberta/encoder/layer.0/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.0/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="148" name="/roberta/encoder/layer.0/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="149" name="Constant_312" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_312" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="150" name="/roberta/encoder/layer.0/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_1, /roberta/encoder/layer.0/attention/self/Gather_1, Constant_312" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="151" name="Constant_316" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_316" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="152" name="/roberta/encoder/layer.0/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Unsqueeze_1, Constant_316" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.0/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="153" name="/roberta/encoder/layer.0/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="154" name="/roberta/encoder/layer.0/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="155" name="/roberta/encoder/layer.0/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Concat, /roberta/encoder/layer.0/attention/self/Constant_2, /roberta/encoder/layer.0/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.0/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="156" name="/roberta/encoder/layer.0/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="157" name="Constant_390" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_390" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="158" name="/roberta/encoder/layer.0/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="159" name="/roberta/encoder/layer.0/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="160" name="/roberta/encoder/layer.0/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="161" name="/roberta/encoder/layer.0/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="162" name="/roberta/Constant_8" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/Constant_8_output_0" /> + </output> + </layer> + <layer id="163" name="Constant_262" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_262" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="164" name="/roberta/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Unsqueeze_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Unsqueeze_3_output_0"> + <dim>-1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="165" name="Constant_264" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2776376460" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_264" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="166" name="/roberta/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Unsqueeze_4" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>-1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/Unsqueeze_4_output_0"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="167" name="/roberta/Cast" type="Convert" version="opset1"> + <data destination_type="f32" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Cast" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/Cast_output_0"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="168" name="/roberta/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32" /> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/Sub_output_0"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="169" name="/roberta/Constant_9" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376468" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/Constant_9_output_0" /> + </output> + </layer> + <layer id="170" name="/roberta/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/Mul_1_output_0"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="171" name="/roberta/encoder/layer.0/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="172" name="/roberta/encoder/layer.0/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="173" name="roberta.encoder.layer.0.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2776376472" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="174" name="onnx::MatMul_1560" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2776379544" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1560" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1560"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="175" name="/roberta/encoder/layer.0/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="176" name="/roberta/encoder/layer.0/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="177" name="/roberta/encoder/layer.0/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.0/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="178" name="/roberta/encoder/layer.0/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="179" name="Constant_336" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_336" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="180" name="/roberta/encoder/layer.0/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_4, /roberta/encoder/layer.0/attention/self/Gather_2, Constant_336" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="181" name="Constant_342" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_342" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="182" name="/roberta/encoder/layer.0/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Unsqueeze_2, Constant_342" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.0/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="183" name="/roberta/encoder/layer.0/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.0/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="184" name="/roberta/encoder/layer.0/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="185" name="Constant_340" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_340" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="186" name="/roberta/encoder/layer.0/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_5, /roberta/encoder/layer.0/attention/self/Gather_3, Constant_340" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="187" name="Constant_344" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_344" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="188" name="/roberta/encoder/layer.0/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Unsqueeze_3, Constant_344" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.0/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="189" name="/roberta/encoder/layer.0/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="190" name="/roberta/encoder/layer.0/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="191" name="/roberta/encoder/layer.0/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Concat_1, /roberta/encoder/layer.0/attention/self/Constant_6, /roberta/encoder/layer.0/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.0/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="192" name="/roberta/encoder/layer.0/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="193" name="Constant_360" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_360" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="194" name="/roberta/encoder/layer.0/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="195" name="/roberta/encoder/layer.0/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="196" name="Constant_398" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_398" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="197" name="/roberta/encoder/layer.0/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="198" name="/roberta/encoder/layer.0/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.0/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="199" name="/roberta/encoder/layer.0/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="200" name="Constant_402" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_402" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="201" name="/roberta/encoder/layer.0/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_13, /roberta/encoder/layer.0/attention/self/Gather_6, Constant_402" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="202" name="Constant_408" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_408" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="203" name="/roberta/encoder/layer.0/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Unsqueeze_6, Constant_408" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.0/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="204" name="/roberta/encoder/layer.0/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.0/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="205" name="/roberta/encoder/layer.0/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="206" name="Constant_406" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_406" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="207" name="/roberta/encoder/layer.0/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_14, /roberta/encoder/layer.0/attention/self/Gather_7, Constant_406" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="208" name="Constant_410" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_410" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="209" name="/roberta/encoder/layer.0/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Unsqueeze_7, Constant_410" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.0/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="210" name="/roberta/encoder/layer.0/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.0/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="211" name="/roberta/encoder/layer.0/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Concat_3, /roberta/encoder/layer.0/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.0/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="212" name="/roberta/encoder/layer.0/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="213" name="onnx::MatMul_1566" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2778738848" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1566" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1566"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="214" name="/roberta/encoder/layer.0/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="215" name="/roberta/encoder/layer.0/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="216" name="/roberta/encoder/layer.0/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="217" name="Constant_428" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_428" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="218" name="/roberta/encoder/layer.0/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="219" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="220" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="221" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="222" name="Constant_433" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_433" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="223" name="/roberta/encoder/layer.0/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="224" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="225" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="226" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="227" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="228" name="roberta.encoder.layer.0.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2781098144" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="229" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="230" name="roberta.encoder.layer.0.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2781101216" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="231" name="/roberta/encoder/layer.0/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="232" name="onnx::MatMul_1567" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="2781104288" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1567" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1567"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="233" name="/roberta/encoder/layer.0/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="234" name="/roberta/encoder/layer.0/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="235" name="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="236" name="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="237" name="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="238" name="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="239" name="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="240" name="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="241" name="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="242" name="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="243" name="onnx::MatMul_1568" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="2790541480" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1568" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1568"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="244" name="/roberta/encoder/layer.0/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="245" name="/roberta/encoder/layer.0/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="246" name="/roberta/encoder/layer.0/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="247" name="Constant_454" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_454" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="248" name="/roberta/encoder/layer.0/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="249" name="/roberta/encoder/layer.0/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="250" name="/roberta/encoder/layer.0/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="251" name="/roberta/encoder/layer.0/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="252" name="Constant_459" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_459" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="253" name="/roberta/encoder/layer.0/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="254" name="/roberta/encoder/layer.0/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="255" name="/roberta/encoder/layer.0/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="256" name="/roberta/encoder/layer.0/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="257" name="/roberta/encoder/layer.0/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="258" name="roberta.encoder.layer.0.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2799978664" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="259" name="/roberta/encoder/layer.0/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="260" name="roberta.encoder.layer.0.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2799981736" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.0.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.0.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="261" name="/roberta/encoder/layer.0/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.0/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.0/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="262" name="onnx::MatMul_1569" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2799984808" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1569" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1569"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="263" name="/roberta/encoder/layer.1/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="264" name="/roberta/encoder/layer.1/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="265" name="/roberta/encoder/layer.1/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.1/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="266" name="/roberta/encoder/layer.1/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="267" name="Constant_529" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_529" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="268" name="/roberta/encoder/layer.1/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_8, /roberta/encoder/layer.1/attention/self/Gather_4, Constant_529" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="269" name="Constant_535" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_535" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="270" name="/roberta/encoder/layer.1/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Unsqueeze_4, Constant_535" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.1/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="271" name="/roberta/encoder/layer.1/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.1/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="272" name="/roberta/encoder/layer.1/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="273" name="Constant_533" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_533" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="274" name="/roberta/encoder/layer.1/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_9, /roberta/encoder/layer.1/attention/self/Gather_5, Constant_533" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="275" name="Constant_537" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_537" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="276" name="/roberta/encoder/layer.1/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Unsqueeze_5, Constant_537" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.1/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="277" name="/roberta/encoder/layer.1/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="278" name="/roberta/encoder/layer.1/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="279" name="/roberta/encoder/layer.1/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Concat_2, /roberta/encoder/layer.1/attention/self/Constant_10, /roberta/encoder/layer.1/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.1/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="280" name="/roberta/encoder/layer.1/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="281" name="Constant_553" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_553" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="282" name="/roberta/encoder/layer.1/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="283" name="roberta.encoder.layer.1.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2802344104" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="284" name="onnx::MatMul_1570" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2802347176" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1570" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1570"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="285" name="/roberta/encoder/layer.1/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="286" name="/roberta/encoder/layer.1/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="287" name="/roberta/encoder/layer.1/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.1/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="288" name="/roberta/encoder/layer.1/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="289" name="Constant_473" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_473" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="290" name="/roberta/encoder/layer.1/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant, /roberta/encoder/layer.1/attention/self/Gather, Constant_473" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="291" name="Constant_479" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_479" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="292" name="/roberta/encoder/layer.1/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Unsqueeze, Constant_479" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.1/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="293" name="/roberta/encoder/layer.1/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.1/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="294" name="/roberta/encoder/layer.1/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="295" name="Constant_477" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_477" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="296" name="/roberta/encoder/layer.1/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_1, /roberta/encoder/layer.1/attention/self/Gather_1, Constant_477" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="297" name="Constant_481" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_481" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="298" name="/roberta/encoder/layer.1/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Unsqueeze_1, Constant_481" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.1/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="299" name="/roberta/encoder/layer.1/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="300" name="/roberta/encoder/layer.1/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="301" name="/roberta/encoder/layer.1/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Concat, /roberta/encoder/layer.1/attention/self/Constant_2, /roberta/encoder/layer.1/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.1/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="302" name="/roberta/encoder/layer.1/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="303" name="Constant_555" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_555" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="304" name="/roberta/encoder/layer.1/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="305" name="/roberta/encoder/layer.1/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="306" name="/roberta/encoder/layer.1/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="307" name="/roberta/encoder/layer.1/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="308" name="/roberta/encoder/layer.1/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="309" name="/roberta/encoder/layer.1/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="310" name="roberta.encoder.layer.1.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2804706472" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="311" name="onnx::MatMul_1573" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2804709544" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1573" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1573"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="312" name="/roberta/encoder/layer.1/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="313" name="/roberta/encoder/layer.1/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="314" name="/roberta/encoder/layer.1/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.1/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="315" name="/roberta/encoder/layer.1/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="316" name="Constant_501" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_501" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="317" name="/roberta/encoder/layer.1/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_4, /roberta/encoder/layer.1/attention/self/Gather_2, Constant_501" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="318" name="Constant_507" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_507" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="319" name="/roberta/encoder/layer.1/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Unsqueeze_2, Constant_507" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.1/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="320" name="/roberta/encoder/layer.1/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.1/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="321" name="/roberta/encoder/layer.1/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="322" name="Constant_505" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_505" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="323" name="/roberta/encoder/layer.1/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_5, /roberta/encoder/layer.1/attention/self/Gather_3, Constant_505" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="324" name="Constant_509" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_509" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="325" name="/roberta/encoder/layer.1/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Unsqueeze_3, Constant_509" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.1/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="326" name="/roberta/encoder/layer.1/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="327" name="/roberta/encoder/layer.1/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="328" name="/roberta/encoder/layer.1/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Concat_1, /roberta/encoder/layer.1/attention/self/Constant_6, /roberta/encoder/layer.1/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.1/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="329" name="/roberta/encoder/layer.1/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="330" name="Constant_525" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_525" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="331" name="/roberta/encoder/layer.1/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="332" name="/roberta/encoder/layer.1/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="333" name="Constant_563" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_563" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="334" name="/roberta/encoder/layer.1/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="335" name="/roberta/encoder/layer.1/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.1/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="336" name="/roberta/encoder/layer.1/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="337" name="Constant_567" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_567" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="338" name="/roberta/encoder/layer.1/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_13, /roberta/encoder/layer.1/attention/self/Gather_6, Constant_567" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="339" name="Constant_573" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_573" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="340" name="/roberta/encoder/layer.1/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Unsqueeze_6, Constant_573" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.1/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="341" name="/roberta/encoder/layer.1/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.1/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="342" name="/roberta/encoder/layer.1/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="343" name="Constant_571" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_571" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="344" name="/roberta/encoder/layer.1/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_14, /roberta/encoder/layer.1/attention/self/Gather_7, Constant_571" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="345" name="Constant_575" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_575" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="346" name="/roberta/encoder/layer.1/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Unsqueeze_7, Constant_575" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.1/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="347" name="/roberta/encoder/layer.1/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.1/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="348" name="/roberta/encoder/layer.1/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Concat_3, /roberta/encoder/layer.1/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.1/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="349" name="/roberta/encoder/layer.1/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="350" name="onnx::MatMul_1579" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2807068840" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1579" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1579"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="351" name="/roberta/encoder/layer.1/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="352" name="/roberta/encoder/layer.1/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="353" name="/roberta/encoder/layer.1/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="354" name="Constant_593" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_593" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="355" name="/roberta/encoder/layer.1/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="356" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="357" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="358" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="359" name="Constant_598" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_598" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="360" name="/roberta/encoder/layer.1/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="361" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="362" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="363" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="364" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="365" name="roberta.encoder.layer.1.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2809428136" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="366" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="367" name="roberta.encoder.layer.1.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2809431208" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="368" name="/roberta/encoder/layer.1/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="369" name="onnx::MatMul_1580" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="2809434280" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1580" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1580"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="370" name="/roberta/encoder/layer.1/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="371" name="/roberta/encoder/layer.1/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="372" name="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="373" name="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="374" name="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="375" name="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="376" name="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="377" name="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="378" name="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="379" name="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="380" name="onnx::MatMul_1581" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="2818871464" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1581" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1581"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="381" name="/roberta/encoder/layer.1/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="382" name="/roberta/encoder/layer.1/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="383" name="/roberta/encoder/layer.1/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="384" name="Constant_619" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_619" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="385" name="/roberta/encoder/layer.1/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="386" name="/roberta/encoder/layer.1/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="387" name="/roberta/encoder/layer.1/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="388" name="/roberta/encoder/layer.1/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="389" name="Constant_624" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_624" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="390" name="/roberta/encoder/layer.1/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="391" name="/roberta/encoder/layer.1/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="392" name="/roberta/encoder/layer.1/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="393" name="/roberta/encoder/layer.1/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="394" name="/roberta/encoder/layer.1/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="395" name="roberta.encoder.layer.1.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2828308648" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="396" name="/roberta/encoder/layer.1/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="397" name="roberta.encoder.layer.1.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2828311720" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.1.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.1.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="398" name="/roberta/encoder/layer.1/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.1/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.1/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="399" name="onnx::MatMul_1582" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2828314792" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1582" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1582"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="400" name="/roberta/encoder/layer.2/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="401" name="/roberta/encoder/layer.2/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="402" name="/roberta/encoder/layer.2/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.2/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="403" name="/roberta/encoder/layer.2/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="404" name="Constant_694" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_694" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="405" name="/roberta/encoder/layer.2/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_8, /roberta/encoder/layer.2/attention/self/Gather_4, Constant_694" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="406" name="Constant_700" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_700" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="407" name="/roberta/encoder/layer.2/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Unsqueeze_4, Constant_700" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.2/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="408" name="/roberta/encoder/layer.2/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.2/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="409" name="/roberta/encoder/layer.2/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="410" name="Constant_698" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_698" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="411" name="/roberta/encoder/layer.2/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_9, /roberta/encoder/layer.2/attention/self/Gather_5, Constant_698" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="412" name="Constant_702" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_702" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="413" name="/roberta/encoder/layer.2/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Unsqueeze_5, Constant_702" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.2/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="414" name="/roberta/encoder/layer.2/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="415" name="/roberta/encoder/layer.2/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="416" name="/roberta/encoder/layer.2/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Concat_2, /roberta/encoder/layer.2/attention/self/Constant_10, /roberta/encoder/layer.2/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.2/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="417" name="/roberta/encoder/layer.2/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="418" name="Constant_718" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_718" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="419" name="/roberta/encoder/layer.2/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="420" name="roberta.encoder.layer.2.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2830674088" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="421" name="onnx::MatMul_1583" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2830677160" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1583" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1583"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="422" name="/roberta/encoder/layer.2/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="423" name="/roberta/encoder/layer.2/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="424" name="/roberta/encoder/layer.2/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.2/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="425" name="/roberta/encoder/layer.2/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="426" name="Constant_638" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_638" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="427" name="/roberta/encoder/layer.2/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant, /roberta/encoder/layer.2/attention/self/Gather, Constant_638" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="428" name="Constant_644" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_644" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="429" name="/roberta/encoder/layer.2/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Unsqueeze, Constant_644" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.2/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="430" name="/roberta/encoder/layer.2/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.2/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="431" name="/roberta/encoder/layer.2/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="432" name="Constant_642" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_642" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="433" name="/roberta/encoder/layer.2/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_1, /roberta/encoder/layer.2/attention/self/Gather_1, Constant_642" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="434" name="Constant_646" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_646" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="435" name="/roberta/encoder/layer.2/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Unsqueeze_1, Constant_646" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.2/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="436" name="/roberta/encoder/layer.2/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="437" name="/roberta/encoder/layer.2/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="438" name="/roberta/encoder/layer.2/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Concat, /roberta/encoder/layer.2/attention/self/Constant_2, /roberta/encoder/layer.2/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.2/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="439" name="/roberta/encoder/layer.2/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="440" name="Constant_720" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_720" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="441" name="/roberta/encoder/layer.2/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="442" name="/roberta/encoder/layer.2/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="443" name="/roberta/encoder/layer.2/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="444" name="/roberta/encoder/layer.2/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="445" name="/roberta/encoder/layer.2/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="446" name="/roberta/encoder/layer.2/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="447" name="roberta.encoder.layer.2.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2833036456" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="448" name="onnx::MatMul_1586" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2833039528" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1586" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1586"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="449" name="/roberta/encoder/layer.2/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="450" name="/roberta/encoder/layer.2/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="451" name="/roberta/encoder/layer.2/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.2/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="452" name="/roberta/encoder/layer.2/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="453" name="Constant_666" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_666" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="454" name="/roberta/encoder/layer.2/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_4, /roberta/encoder/layer.2/attention/self/Gather_2, Constant_666" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="455" name="Constant_672" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_672" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="456" name="/roberta/encoder/layer.2/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Unsqueeze_2, Constant_672" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.2/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="457" name="/roberta/encoder/layer.2/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.2/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="458" name="/roberta/encoder/layer.2/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="459" name="Constant_670" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_670" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="460" name="/roberta/encoder/layer.2/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_5, /roberta/encoder/layer.2/attention/self/Gather_3, Constant_670" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="461" name="Constant_674" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_674" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="462" name="/roberta/encoder/layer.2/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Unsqueeze_3, Constant_674" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.2/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="463" name="/roberta/encoder/layer.2/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="464" name="/roberta/encoder/layer.2/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="465" name="/roberta/encoder/layer.2/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Concat_1, /roberta/encoder/layer.2/attention/self/Constant_6, /roberta/encoder/layer.2/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.2/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="466" name="/roberta/encoder/layer.2/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="467" name="Constant_690" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_690" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="468" name="/roberta/encoder/layer.2/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="469" name="/roberta/encoder/layer.2/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="470" name="Constant_728" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_728" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="471" name="/roberta/encoder/layer.2/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="472" name="/roberta/encoder/layer.2/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.2/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="473" name="/roberta/encoder/layer.2/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="474" name="Constant_732" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_732" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="475" name="/roberta/encoder/layer.2/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_13, /roberta/encoder/layer.2/attention/self/Gather_6, Constant_732" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="476" name="Constant_738" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_738" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="477" name="/roberta/encoder/layer.2/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Unsqueeze_6, Constant_738" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.2/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="478" name="/roberta/encoder/layer.2/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.2/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="479" name="/roberta/encoder/layer.2/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="480" name="Constant_736" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_736" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="481" name="/roberta/encoder/layer.2/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_14, /roberta/encoder/layer.2/attention/self/Gather_7, Constant_736" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="482" name="Constant_740" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_740" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="483" name="/roberta/encoder/layer.2/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Unsqueeze_7, Constant_740" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.2/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="484" name="/roberta/encoder/layer.2/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.2/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="485" name="/roberta/encoder/layer.2/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Concat_3, /roberta/encoder/layer.2/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.2/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="486" name="/roberta/encoder/layer.2/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="487" name="onnx::MatMul_1592" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2835398824" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1592" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1592"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="488" name="/roberta/encoder/layer.2/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="489" name="/roberta/encoder/layer.2/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="490" name="/roberta/encoder/layer.2/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="491" name="Constant_758" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_758" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="492" name="/roberta/encoder/layer.2/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="493" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="494" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="495" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="496" name="Constant_763" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_763" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="497" name="/roberta/encoder/layer.2/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="498" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="499" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="500" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="501" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="502" name="roberta.encoder.layer.2.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2837758120" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="503" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="504" name="roberta.encoder.layer.2.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2837761192" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="505" name="/roberta/encoder/layer.2/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="506" name="onnx::MatMul_1593" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="2837764264" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1593" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1593"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="507" name="/roberta/encoder/layer.2/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="508" name="/roberta/encoder/layer.2/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="509" name="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="510" name="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="511" name="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="512" name="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="513" name="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="514" name="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="515" name="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="516" name="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="517" name="onnx::MatMul_1594" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="2847201448" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1594" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1594"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="518" name="/roberta/encoder/layer.2/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="519" name="/roberta/encoder/layer.2/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="520" name="/roberta/encoder/layer.2/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="521" name="Constant_784" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_784" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="522" name="/roberta/encoder/layer.2/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="523" name="/roberta/encoder/layer.2/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="524" name="/roberta/encoder/layer.2/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="525" name="/roberta/encoder/layer.2/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="526" name="Constant_789" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_789" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="527" name="/roberta/encoder/layer.2/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="528" name="/roberta/encoder/layer.2/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="529" name="/roberta/encoder/layer.2/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="530" name="/roberta/encoder/layer.2/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="531" name="/roberta/encoder/layer.2/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="532" name="roberta.encoder.layer.2.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2856638632" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="533" name="/roberta/encoder/layer.2/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="534" name="roberta.encoder.layer.2.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2856641704" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.2.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.2.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="535" name="/roberta/encoder/layer.2/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.2/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.2/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="536" name="onnx::MatMul_1595" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2856644776" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1595" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1595"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="537" name="/roberta/encoder/layer.3/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="538" name="/roberta/encoder/layer.3/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="539" name="/roberta/encoder/layer.3/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.3/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="540" name="/roberta/encoder/layer.3/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="541" name="Constant_859" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_859" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="542" name="/roberta/encoder/layer.3/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_8, /roberta/encoder/layer.3/attention/self/Gather_4, Constant_859" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="543" name="Constant_865" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_865" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="544" name="/roberta/encoder/layer.3/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Unsqueeze_4, Constant_865" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.3/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="545" name="/roberta/encoder/layer.3/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.3/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="546" name="/roberta/encoder/layer.3/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="547" name="Constant_863" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_863" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="548" name="/roberta/encoder/layer.3/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_9, /roberta/encoder/layer.3/attention/self/Gather_5, Constant_863" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="549" name="Constant_867" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_867" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="550" name="/roberta/encoder/layer.3/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Unsqueeze_5, Constant_867" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.3/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="551" name="/roberta/encoder/layer.3/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="552" name="/roberta/encoder/layer.3/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="553" name="/roberta/encoder/layer.3/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Concat_2, /roberta/encoder/layer.3/attention/self/Constant_10, /roberta/encoder/layer.3/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.3/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="554" name="/roberta/encoder/layer.3/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="555" name="Constant_883" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_883" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="556" name="/roberta/encoder/layer.3/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="557" name="roberta.encoder.layer.3.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2859004072" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="558" name="onnx::MatMul_1596" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2859007144" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1596" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1596"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="559" name="/roberta/encoder/layer.3/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="560" name="/roberta/encoder/layer.3/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="561" name="/roberta/encoder/layer.3/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.3/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="562" name="/roberta/encoder/layer.3/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="563" name="Constant_803" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_803" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="564" name="/roberta/encoder/layer.3/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant, /roberta/encoder/layer.3/attention/self/Gather, Constant_803" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="565" name="Constant_809" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_809" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="566" name="/roberta/encoder/layer.3/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Unsqueeze, Constant_809" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.3/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="567" name="/roberta/encoder/layer.3/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.3/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="568" name="/roberta/encoder/layer.3/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="569" name="Constant_807" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_807" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="570" name="/roberta/encoder/layer.3/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_1, /roberta/encoder/layer.3/attention/self/Gather_1, Constant_807" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="571" name="Constant_811" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_811" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="572" name="/roberta/encoder/layer.3/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Unsqueeze_1, Constant_811" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.3/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="573" name="/roberta/encoder/layer.3/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="574" name="/roberta/encoder/layer.3/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="575" name="/roberta/encoder/layer.3/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Concat, /roberta/encoder/layer.3/attention/self/Constant_2, /roberta/encoder/layer.3/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.3/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="576" name="/roberta/encoder/layer.3/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="577" name="Constant_885" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_885" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="578" name="/roberta/encoder/layer.3/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="579" name="/roberta/encoder/layer.3/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="580" name="/roberta/encoder/layer.3/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="581" name="/roberta/encoder/layer.3/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="582" name="/roberta/encoder/layer.3/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="583" name="/roberta/encoder/layer.3/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="584" name="roberta.encoder.layer.3.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2861366440" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="585" name="onnx::MatMul_1599" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2861369512" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1599" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1599"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="586" name="/roberta/encoder/layer.3/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="587" name="/roberta/encoder/layer.3/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="588" name="/roberta/encoder/layer.3/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.3/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="589" name="/roberta/encoder/layer.3/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="590" name="Constant_831" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_831" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="591" name="/roberta/encoder/layer.3/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_4, /roberta/encoder/layer.3/attention/self/Gather_2, Constant_831" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="592" name="Constant_837" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_837" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="593" name="/roberta/encoder/layer.3/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Unsqueeze_2, Constant_837" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.3/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="594" name="/roberta/encoder/layer.3/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.3/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="595" name="/roberta/encoder/layer.3/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="596" name="Constant_835" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_835" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="597" name="/roberta/encoder/layer.3/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_5, /roberta/encoder/layer.3/attention/self/Gather_3, Constant_835" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="598" name="Constant_839" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_839" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="599" name="/roberta/encoder/layer.3/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Unsqueeze_3, Constant_839" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.3/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="600" name="/roberta/encoder/layer.3/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="601" name="/roberta/encoder/layer.3/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="602" name="/roberta/encoder/layer.3/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Concat_1, /roberta/encoder/layer.3/attention/self/Constant_6, /roberta/encoder/layer.3/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.3/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="603" name="/roberta/encoder/layer.3/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="604" name="Constant_855" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_855" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="605" name="/roberta/encoder/layer.3/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="606" name="/roberta/encoder/layer.3/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="607" name="Constant_893" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_893" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="608" name="/roberta/encoder/layer.3/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="609" name="/roberta/encoder/layer.3/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.3/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="610" name="/roberta/encoder/layer.3/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="611" name="Constant_897" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_897" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="612" name="/roberta/encoder/layer.3/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_13, /roberta/encoder/layer.3/attention/self/Gather_6, Constant_897" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="613" name="Constant_903" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_903" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="614" name="/roberta/encoder/layer.3/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Unsqueeze_6, Constant_903" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.3/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="615" name="/roberta/encoder/layer.3/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.3/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="616" name="/roberta/encoder/layer.3/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="617" name="Constant_901" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_901" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="618" name="/roberta/encoder/layer.3/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_14, /roberta/encoder/layer.3/attention/self/Gather_7, Constant_901" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="619" name="Constant_905" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_905" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="620" name="/roberta/encoder/layer.3/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Unsqueeze_7, Constant_905" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.3/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="621" name="/roberta/encoder/layer.3/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.3/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="622" name="/roberta/encoder/layer.3/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Concat_3, /roberta/encoder/layer.3/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.3/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="623" name="/roberta/encoder/layer.3/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="624" name="onnx::MatMul_1605" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2863728808" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1605" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1605"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="625" name="/roberta/encoder/layer.3/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="626" name="/roberta/encoder/layer.3/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="627" name="/roberta/encoder/layer.3/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="628" name="Constant_923" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_923" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="629" name="/roberta/encoder/layer.3/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="630" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="631" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="632" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="633" name="Constant_928" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_928" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="634" name="/roberta/encoder/layer.3/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="635" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="636" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="637" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="638" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="639" name="roberta.encoder.layer.3.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2866088104" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="640" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="641" name="roberta.encoder.layer.3.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2866091176" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="642" name="/roberta/encoder/layer.3/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="643" name="onnx::MatMul_1606" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="2866094248" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1606" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1606"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="644" name="/roberta/encoder/layer.3/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="645" name="/roberta/encoder/layer.3/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="646" name="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="647" name="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="648" name="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="649" name="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="650" name="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="651" name="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="652" name="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="653" name="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="654" name="onnx::MatMul_1607" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="2875531432" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1607" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1607"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="655" name="/roberta/encoder/layer.3/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="656" name="/roberta/encoder/layer.3/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="657" name="/roberta/encoder/layer.3/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="658" name="Constant_949" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_949" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="659" name="/roberta/encoder/layer.3/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="660" name="/roberta/encoder/layer.3/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="661" name="/roberta/encoder/layer.3/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="662" name="/roberta/encoder/layer.3/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="663" name="Constant_954" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_954" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="664" name="/roberta/encoder/layer.3/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="665" name="/roberta/encoder/layer.3/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="666" name="/roberta/encoder/layer.3/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="667" name="/roberta/encoder/layer.3/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="668" name="/roberta/encoder/layer.3/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="669" name="roberta.encoder.layer.3.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2884968616" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="670" name="/roberta/encoder/layer.3/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="671" name="roberta.encoder.layer.3.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2884971688" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.3.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.3.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="672" name="/roberta/encoder/layer.3/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.3/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.3/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="673" name="onnx::MatMul_1608" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2884974760" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1608" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1608"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="674" name="/roberta/encoder/layer.4/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="675" name="/roberta/encoder/layer.4/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="676" name="/roberta/encoder/layer.4/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.4/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="677" name="/roberta/encoder/layer.4/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="678" name="Constant_1024" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1024" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="679" name="/roberta/encoder/layer.4/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_8, /roberta/encoder/layer.4/attention/self/Gather_4, Constant_1024" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="680" name="Constant_1030" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1030" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="681" name="/roberta/encoder/layer.4/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Unsqueeze_4, Constant_1030" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.4/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="682" name="/roberta/encoder/layer.4/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.4/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="683" name="/roberta/encoder/layer.4/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="684" name="Constant_1028" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1028" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="685" name="/roberta/encoder/layer.4/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_9, /roberta/encoder/layer.4/attention/self/Gather_5, Constant_1028" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="686" name="Constant_1032" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1032" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="687" name="/roberta/encoder/layer.4/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Unsqueeze_5, Constant_1032" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.4/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="688" name="/roberta/encoder/layer.4/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="689" name="/roberta/encoder/layer.4/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="690" name="/roberta/encoder/layer.4/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Concat_2, /roberta/encoder/layer.4/attention/self/Constant_10, /roberta/encoder/layer.4/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.4/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="691" name="/roberta/encoder/layer.4/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="692" name="Constant_1048" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1048" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="693" name="/roberta/encoder/layer.4/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="694" name="roberta.encoder.layer.4.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2887334056" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="695" name="onnx::MatMul_1609" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2887337128" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1609" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1609"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="696" name="/roberta/encoder/layer.4/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="697" name="/roberta/encoder/layer.4/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="698" name="/roberta/encoder/layer.4/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.4/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="699" name="/roberta/encoder/layer.4/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="700" name="Constant_968" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_968" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="701" name="/roberta/encoder/layer.4/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant, /roberta/encoder/layer.4/attention/self/Gather, Constant_968" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="702" name="Constant_974" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_974" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="703" name="/roberta/encoder/layer.4/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Unsqueeze, Constant_974" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.4/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="704" name="/roberta/encoder/layer.4/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.4/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="705" name="/roberta/encoder/layer.4/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="706" name="Constant_972" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_972" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="707" name="/roberta/encoder/layer.4/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_1, /roberta/encoder/layer.4/attention/self/Gather_1, Constant_972" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="708" name="Constant_976" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_976" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="709" name="/roberta/encoder/layer.4/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Unsqueeze_1, Constant_976" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.4/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="710" name="/roberta/encoder/layer.4/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="711" name="/roberta/encoder/layer.4/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="712" name="/roberta/encoder/layer.4/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Concat, /roberta/encoder/layer.4/attention/self/Constant_2, /roberta/encoder/layer.4/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.4/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="713" name="/roberta/encoder/layer.4/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="714" name="Constant_1050" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1050" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="715" name="/roberta/encoder/layer.4/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="716" name="/roberta/encoder/layer.4/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="717" name="/roberta/encoder/layer.4/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="718" name="/roberta/encoder/layer.4/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="719" name="/roberta/encoder/layer.4/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="720" name="/roberta/encoder/layer.4/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="721" name="roberta.encoder.layer.4.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2889696424" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="722" name="onnx::MatMul_1612" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2889699496" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1612" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1612"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="723" name="/roberta/encoder/layer.4/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="724" name="/roberta/encoder/layer.4/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="725" name="/roberta/encoder/layer.4/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.4/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="726" name="/roberta/encoder/layer.4/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="727" name="Constant_996" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_996" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="728" name="/roberta/encoder/layer.4/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_4, /roberta/encoder/layer.4/attention/self/Gather_2, Constant_996" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="729" name="Constant_1002" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1002" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="730" name="/roberta/encoder/layer.4/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Unsqueeze_2, Constant_1002" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.4/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="731" name="/roberta/encoder/layer.4/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.4/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="732" name="/roberta/encoder/layer.4/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="733" name="Constant_1000" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1000" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="734" name="/roberta/encoder/layer.4/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_5, /roberta/encoder/layer.4/attention/self/Gather_3, Constant_1000" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="735" name="Constant_1004" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1004" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="736" name="/roberta/encoder/layer.4/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Unsqueeze_3, Constant_1004" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.4/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="737" name="/roberta/encoder/layer.4/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="738" name="/roberta/encoder/layer.4/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="739" name="/roberta/encoder/layer.4/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Concat_1, /roberta/encoder/layer.4/attention/self/Constant_6, /roberta/encoder/layer.4/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.4/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="740" name="/roberta/encoder/layer.4/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="741" name="Constant_1020" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1020" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="742" name="/roberta/encoder/layer.4/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="743" name="/roberta/encoder/layer.4/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="744" name="Constant_1058" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1058" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="745" name="/roberta/encoder/layer.4/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="746" name="/roberta/encoder/layer.4/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.4/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="747" name="/roberta/encoder/layer.4/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="748" name="Constant_1062" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1062" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="749" name="/roberta/encoder/layer.4/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_13, /roberta/encoder/layer.4/attention/self/Gather_6, Constant_1062" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="750" name="Constant_1068" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1068" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="751" name="/roberta/encoder/layer.4/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Unsqueeze_6, Constant_1068" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.4/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="752" name="/roberta/encoder/layer.4/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.4/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="753" name="/roberta/encoder/layer.4/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="754" name="Constant_1066" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1066" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="755" name="/roberta/encoder/layer.4/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_14, /roberta/encoder/layer.4/attention/self/Gather_7, Constant_1066" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="756" name="Constant_1070" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1070" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="757" name="/roberta/encoder/layer.4/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Unsqueeze_7, Constant_1070" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.4/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="758" name="/roberta/encoder/layer.4/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.4/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="759" name="/roberta/encoder/layer.4/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Concat_3, /roberta/encoder/layer.4/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.4/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="760" name="/roberta/encoder/layer.4/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="761" name="onnx::MatMul_1618" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2892058792" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1618" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1618"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="762" name="/roberta/encoder/layer.4/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="763" name="/roberta/encoder/layer.4/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="764" name="/roberta/encoder/layer.4/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="765" name="Constant_1088" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1088" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="766" name="/roberta/encoder/layer.4/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="767" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="768" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="769" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="770" name="Constant_1093" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1093" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="771" name="/roberta/encoder/layer.4/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="772" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="773" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="774" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="775" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="776" name="roberta.encoder.layer.4.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2894418088" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="777" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="778" name="roberta.encoder.layer.4.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2894421160" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="779" name="/roberta/encoder/layer.4/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="780" name="onnx::MatMul_1619" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="2894424232" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1619" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1619"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="781" name="/roberta/encoder/layer.4/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="782" name="/roberta/encoder/layer.4/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="783" name="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="784" name="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="785" name="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="786" name="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="787" name="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="788" name="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="789" name="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="790" name="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="791" name="onnx::MatMul_1620" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="2903861416" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1620" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1620"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="792" name="/roberta/encoder/layer.4/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="793" name="/roberta/encoder/layer.4/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="794" name="/roberta/encoder/layer.4/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="795" name="Constant_1114" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1114" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="796" name="/roberta/encoder/layer.4/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="797" name="/roberta/encoder/layer.4/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="798" name="/roberta/encoder/layer.4/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="799" name="/roberta/encoder/layer.4/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="800" name="Constant_1119" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1119" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="801" name="/roberta/encoder/layer.4/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="802" name="/roberta/encoder/layer.4/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="803" name="/roberta/encoder/layer.4/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="804" name="/roberta/encoder/layer.4/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="805" name="/roberta/encoder/layer.4/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="806" name="roberta.encoder.layer.4.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2913298600" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="807" name="/roberta/encoder/layer.4/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="808" name="roberta.encoder.layer.4.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2913301672" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.4.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.4.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="809" name="/roberta/encoder/layer.4/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.4/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.4/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="810" name="onnx::MatMul_1621" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2913304744" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1621" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1621"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="811" name="/roberta/encoder/layer.5/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="812" name="/roberta/encoder/layer.5/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="813" name="/roberta/encoder/layer.5/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.5/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="814" name="/roberta/encoder/layer.5/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="815" name="Constant_1189" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1189" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="816" name="/roberta/encoder/layer.5/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_8, /roberta/encoder/layer.5/attention/self/Gather_4, Constant_1189" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="817" name="Constant_1195" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1195" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="818" name="/roberta/encoder/layer.5/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Unsqueeze_4, Constant_1195" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.5/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="819" name="/roberta/encoder/layer.5/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.5/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="820" name="/roberta/encoder/layer.5/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="821" name="Constant_1193" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1193" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="822" name="/roberta/encoder/layer.5/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_9, /roberta/encoder/layer.5/attention/self/Gather_5, Constant_1193" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="823" name="Constant_1197" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1197" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="824" name="/roberta/encoder/layer.5/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Unsqueeze_5, Constant_1197" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.5/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="825" name="/roberta/encoder/layer.5/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="826" name="/roberta/encoder/layer.5/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="827" name="/roberta/encoder/layer.5/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Concat_2, /roberta/encoder/layer.5/attention/self/Constant_10, /roberta/encoder/layer.5/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.5/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="828" name="/roberta/encoder/layer.5/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="829" name="Constant_1213" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1213" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="830" name="/roberta/encoder/layer.5/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="831" name="roberta.encoder.layer.5.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2915664040" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="832" name="onnx::MatMul_1622" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2915667112" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1622" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1622"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="833" name="/roberta/encoder/layer.5/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="834" name="/roberta/encoder/layer.5/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="835" name="/roberta/encoder/layer.5/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.5/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="836" name="/roberta/encoder/layer.5/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="837" name="Constant_1133" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1133" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="838" name="/roberta/encoder/layer.5/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant, /roberta/encoder/layer.5/attention/self/Gather, Constant_1133" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="839" name="Constant_1139" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1139" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="840" name="/roberta/encoder/layer.5/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Unsqueeze, Constant_1139" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.5/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="841" name="/roberta/encoder/layer.5/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.5/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="842" name="/roberta/encoder/layer.5/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="843" name="Constant_1137" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1137" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="844" name="/roberta/encoder/layer.5/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_1, /roberta/encoder/layer.5/attention/self/Gather_1, Constant_1137" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="845" name="Constant_1141" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1141" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="846" name="/roberta/encoder/layer.5/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Unsqueeze_1, Constant_1141" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.5/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="847" name="/roberta/encoder/layer.5/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="848" name="/roberta/encoder/layer.5/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="849" name="/roberta/encoder/layer.5/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Concat, /roberta/encoder/layer.5/attention/self/Constant_2, /roberta/encoder/layer.5/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.5/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="850" name="/roberta/encoder/layer.5/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="851" name="Constant_1215" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1215" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="852" name="/roberta/encoder/layer.5/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="853" name="/roberta/encoder/layer.5/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="854" name="/roberta/encoder/layer.5/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="855" name="/roberta/encoder/layer.5/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="856" name="/roberta/encoder/layer.5/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="857" name="/roberta/encoder/layer.5/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="858" name="roberta.encoder.layer.5.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2918026408" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="859" name="onnx::MatMul_1625" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2918029480" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1625" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1625"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="860" name="/roberta/encoder/layer.5/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="861" name="/roberta/encoder/layer.5/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="862" name="/roberta/encoder/layer.5/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.5/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="863" name="/roberta/encoder/layer.5/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="864" name="Constant_1161" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1161" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="865" name="/roberta/encoder/layer.5/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_4, /roberta/encoder/layer.5/attention/self/Gather_2, Constant_1161" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="866" name="Constant_1167" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1167" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="867" name="/roberta/encoder/layer.5/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Unsqueeze_2, Constant_1167" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.5/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="868" name="/roberta/encoder/layer.5/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.5/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="869" name="/roberta/encoder/layer.5/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="870" name="Constant_1165" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1165" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="871" name="/roberta/encoder/layer.5/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_5, /roberta/encoder/layer.5/attention/self/Gather_3, Constant_1165" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="872" name="Constant_1169" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1169" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="873" name="/roberta/encoder/layer.5/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Unsqueeze_3, Constant_1169" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.5/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="874" name="/roberta/encoder/layer.5/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="875" name="/roberta/encoder/layer.5/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="876" name="/roberta/encoder/layer.5/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Concat_1, /roberta/encoder/layer.5/attention/self/Constant_6, /roberta/encoder/layer.5/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.5/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="877" name="/roberta/encoder/layer.5/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="878" name="Constant_1185" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1185" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="879" name="/roberta/encoder/layer.5/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="880" name="/roberta/encoder/layer.5/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="881" name="Constant_1223" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1223" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="882" name="/roberta/encoder/layer.5/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="883" name="/roberta/encoder/layer.5/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.5/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="884" name="/roberta/encoder/layer.5/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="885" name="Constant_1227" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1227" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="886" name="/roberta/encoder/layer.5/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_13, /roberta/encoder/layer.5/attention/self/Gather_6, Constant_1227" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="887" name="Constant_1233" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1233" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="888" name="/roberta/encoder/layer.5/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Unsqueeze_6, Constant_1233" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.5/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="889" name="/roberta/encoder/layer.5/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.5/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="890" name="/roberta/encoder/layer.5/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="891" name="Constant_1231" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1231" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="892" name="/roberta/encoder/layer.5/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_14, /roberta/encoder/layer.5/attention/self/Gather_7, Constant_1231" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="893" name="Constant_1235" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1235" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="894" name="/roberta/encoder/layer.5/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Unsqueeze_7, Constant_1235" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.5/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="895" name="/roberta/encoder/layer.5/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.5/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="896" name="/roberta/encoder/layer.5/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Concat_3, /roberta/encoder/layer.5/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.5/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="897" name="/roberta/encoder/layer.5/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="898" name="onnx::MatMul_1631" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2920388776" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1631" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1631"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="899" name="/roberta/encoder/layer.5/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="900" name="/roberta/encoder/layer.5/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="901" name="/roberta/encoder/layer.5/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="902" name="Constant_1253" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1253" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="903" name="/roberta/encoder/layer.5/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="904" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="905" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="906" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="907" name="Constant_1258" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1258" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="908" name="/roberta/encoder/layer.5/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="909" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="910" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="911" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="912" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="913" name="roberta.encoder.layer.5.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2922748072" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="914" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="915" name="roberta.encoder.layer.5.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2922751144" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="916" name="/roberta/encoder/layer.5/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="917" name="onnx::MatMul_1632" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="2922754216" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1632" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1632"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="918" name="/roberta/encoder/layer.5/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="919" name="/roberta/encoder/layer.5/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="920" name="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="921" name="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="922" name="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="923" name="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="924" name="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="925" name="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="926" name="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="927" name="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="928" name="onnx::MatMul_1633" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="2932191400" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1633" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1633"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="929" name="/roberta/encoder/layer.5/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="930" name="/roberta/encoder/layer.5/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="931" name="/roberta/encoder/layer.5/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="932" name="Constant_1279" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1279" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="933" name="/roberta/encoder/layer.5/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="934" name="/roberta/encoder/layer.5/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="935" name="/roberta/encoder/layer.5/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="936" name="/roberta/encoder/layer.5/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="937" name="Constant_1284" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1284" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="938" name="/roberta/encoder/layer.5/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="939" name="/roberta/encoder/layer.5/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="940" name="/roberta/encoder/layer.5/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="941" name="/roberta/encoder/layer.5/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="942" name="/roberta/encoder/layer.5/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="943" name="roberta.encoder.layer.5.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2941628584" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="944" name="/roberta/encoder/layer.5/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="945" name="roberta.encoder.layer.5.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2941631656" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.5.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.5.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="946" name="/roberta/encoder/layer.5/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.5/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.5/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="947" name="onnx::MatMul_1634" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2941634728" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1634" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1634"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="948" name="/roberta/encoder/layer.6/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="949" name="/roberta/encoder/layer.6/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="950" name="/roberta/encoder/layer.6/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.6/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="951" name="/roberta/encoder/layer.6/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="952" name="Constant_1354" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1354" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="953" name="/roberta/encoder/layer.6/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_8, /roberta/encoder/layer.6/attention/self/Gather_4, Constant_1354" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="954" name="Constant_1360" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1360" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="955" name="/roberta/encoder/layer.6/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Unsqueeze_4, Constant_1360" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.6/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="956" name="/roberta/encoder/layer.6/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.6/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="957" name="/roberta/encoder/layer.6/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="958" name="Constant_1358" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1358" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="959" name="/roberta/encoder/layer.6/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_9, /roberta/encoder/layer.6/attention/self/Gather_5, Constant_1358" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="960" name="Constant_1362" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1362" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="961" name="/roberta/encoder/layer.6/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Unsqueeze_5, Constant_1362" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.6/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="962" name="/roberta/encoder/layer.6/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="963" name="/roberta/encoder/layer.6/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="964" name="/roberta/encoder/layer.6/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Concat_2, /roberta/encoder/layer.6/attention/self/Constant_10, /roberta/encoder/layer.6/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.6/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="965" name="/roberta/encoder/layer.6/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="966" name="Constant_1378" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1378" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="967" name="/roberta/encoder/layer.6/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="968" name="roberta.encoder.layer.6.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2943994024" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="969" name="onnx::MatMul_1635" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2943997096" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1635" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1635"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="970" name="/roberta/encoder/layer.6/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="971" name="/roberta/encoder/layer.6/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="972" name="/roberta/encoder/layer.6/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.6/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="973" name="/roberta/encoder/layer.6/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="974" name="Constant_1298" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1298" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="975" name="/roberta/encoder/layer.6/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant, /roberta/encoder/layer.6/attention/self/Gather, Constant_1298" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="976" name="Constant_1304" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1304" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="977" name="/roberta/encoder/layer.6/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Unsqueeze, Constant_1304" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.6/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="978" name="/roberta/encoder/layer.6/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.6/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="979" name="/roberta/encoder/layer.6/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="980" name="Constant_1302" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1302" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="981" name="/roberta/encoder/layer.6/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_1, /roberta/encoder/layer.6/attention/self/Gather_1, Constant_1302" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="982" name="Constant_1306" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1306" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="983" name="/roberta/encoder/layer.6/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Unsqueeze_1, Constant_1306" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.6/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="984" name="/roberta/encoder/layer.6/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="985" name="/roberta/encoder/layer.6/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="986" name="/roberta/encoder/layer.6/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Concat, /roberta/encoder/layer.6/attention/self/Constant_2, /roberta/encoder/layer.6/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.6/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="987" name="/roberta/encoder/layer.6/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="988" name="Constant_1380" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1380" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="989" name="/roberta/encoder/layer.6/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="990" name="/roberta/encoder/layer.6/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="991" name="/roberta/encoder/layer.6/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="992" name="/roberta/encoder/layer.6/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="993" name="/roberta/encoder/layer.6/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="994" name="/roberta/encoder/layer.6/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="995" name="roberta.encoder.layer.6.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2946356392" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="996" name="onnx::MatMul_1638" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2946359464" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1638" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1638"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="997" name="/roberta/encoder/layer.6/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="998" name="/roberta/encoder/layer.6/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="999" name="/roberta/encoder/layer.6/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.6/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1000" name="/roberta/encoder/layer.6/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="1001" name="Constant_1326" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1326" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1002" name="/roberta/encoder/layer.6/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_4, /roberta/encoder/layer.6/attention/self/Gather_2, Constant_1326" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="1003" name="Constant_1332" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1332" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1004" name="/roberta/encoder/layer.6/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Unsqueeze_2, Constant_1332" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.6/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1005" name="/roberta/encoder/layer.6/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.6/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1006" name="/roberta/encoder/layer.6/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="1007" name="Constant_1330" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1330" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1008" name="/roberta/encoder/layer.6/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_5, /roberta/encoder/layer.6/attention/self/Gather_3, Constant_1330" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="1009" name="Constant_1334" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1334" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1010" name="/roberta/encoder/layer.6/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Unsqueeze_3, Constant_1334" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.6/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1011" name="/roberta/encoder/layer.6/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1012" name="/roberta/encoder/layer.6/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1013" name="/roberta/encoder/layer.6/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Concat_1, /roberta/encoder/layer.6/attention/self/Constant_6, /roberta/encoder/layer.6/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.6/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1014" name="/roberta/encoder/layer.6/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1015" name="Constant_1350" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1350" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1016" name="/roberta/encoder/layer.6/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1017" name="/roberta/encoder/layer.6/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1018" name="Constant_1388" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1388" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1019" name="/roberta/encoder/layer.6/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1020" name="/roberta/encoder/layer.6/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.6/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1021" name="/roberta/encoder/layer.6/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="1022" name="Constant_1392" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1392" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1023" name="/roberta/encoder/layer.6/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_13, /roberta/encoder/layer.6/attention/self/Gather_6, Constant_1392" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="1024" name="Constant_1398" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1398" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1025" name="/roberta/encoder/layer.6/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Unsqueeze_6, Constant_1398" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.6/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1026" name="/roberta/encoder/layer.6/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.6/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1027" name="/roberta/encoder/layer.6/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="1028" name="Constant_1396" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1396" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1029" name="/roberta/encoder/layer.6/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_14, /roberta/encoder/layer.6/attention/self/Gather_7, Constant_1396" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="1030" name="Constant_1400" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1400" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1031" name="/roberta/encoder/layer.6/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Unsqueeze_7, Constant_1400" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.6/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1032" name="/roberta/encoder/layer.6/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.6/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1033" name="/roberta/encoder/layer.6/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Concat_3, /roberta/encoder/layer.6/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.6/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1034" name="/roberta/encoder/layer.6/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1035" name="onnx::MatMul_1644" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2948718760" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1644" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1644"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1036" name="/roberta/encoder/layer.6/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1037" name="/roberta/encoder/layer.6/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1038" name="/roberta/encoder/layer.6/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1039" name="Constant_1418" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1418" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1040" name="/roberta/encoder/layer.6/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1041" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1042" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1043" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1044" name="Constant_1423" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1423" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1045" name="/roberta/encoder/layer.6/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1046" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1047" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1048" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1049" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1050" name="roberta.encoder.layer.6.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2951078056" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1051" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1052" name="roberta.encoder.layer.6.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2951081128" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1053" name="/roberta/encoder/layer.6/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1054" name="onnx::MatMul_1645" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="2951084200" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1645" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1645"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1055" name="/roberta/encoder/layer.6/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1056" name="/roberta/encoder/layer.6/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1057" name="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="1058" name="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1059" name="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1060" name="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="1061" name="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1062" name="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1063" name="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="1064" name="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1065" name="onnx::MatMul_1646" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="2960521384" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1646" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1646"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1066" name="/roberta/encoder/layer.6/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1067" name="/roberta/encoder/layer.6/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1068" name="/roberta/encoder/layer.6/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1069" name="Constant_1444" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1444" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1070" name="/roberta/encoder/layer.6/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1071" name="/roberta/encoder/layer.6/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1072" name="/roberta/encoder/layer.6/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1073" name="/roberta/encoder/layer.6/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1074" name="Constant_1449" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1449" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1075" name="/roberta/encoder/layer.6/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1076" name="/roberta/encoder/layer.6/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1077" name="/roberta/encoder/layer.6/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1078" name="/roberta/encoder/layer.6/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1079" name="/roberta/encoder/layer.6/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1080" name="roberta.encoder.layer.6.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2969958568" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1081" name="/roberta/encoder/layer.6/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1082" name="roberta.encoder.layer.6.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2969961640" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.6.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.6.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1083" name="/roberta/encoder/layer.6/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.6/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.6/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1084" name="onnx::MatMul_1647" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2969964712" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1647" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1647"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1085" name="/roberta/encoder/layer.7/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1086" name="/roberta/encoder/layer.7/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1087" name="/roberta/encoder/layer.7/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.7/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1088" name="/roberta/encoder/layer.7/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="1089" name="Constant_1519" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1519" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1090" name="/roberta/encoder/layer.7/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_8, /roberta/encoder/layer.7/attention/self/Gather_4, Constant_1519" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="1091" name="Constant_1525" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1525" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1092" name="/roberta/encoder/layer.7/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Unsqueeze_4, Constant_1525" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.7/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1093" name="/roberta/encoder/layer.7/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.7/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1094" name="/roberta/encoder/layer.7/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="1095" name="Constant_1523" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1523" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1096" name="/roberta/encoder/layer.7/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_9, /roberta/encoder/layer.7/attention/self/Gather_5, Constant_1523" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="1097" name="Constant_1527" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1527" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1098" name="/roberta/encoder/layer.7/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Unsqueeze_5, Constant_1527" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.7/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1099" name="/roberta/encoder/layer.7/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1100" name="/roberta/encoder/layer.7/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1101" name="/roberta/encoder/layer.7/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Concat_2, /roberta/encoder/layer.7/attention/self/Constant_10, /roberta/encoder/layer.7/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.7/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1102" name="/roberta/encoder/layer.7/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1103" name="Constant_1543" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1543" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1104" name="/roberta/encoder/layer.7/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1105" name="roberta.encoder.layer.7.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2972324008" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1106" name="onnx::MatMul_1648" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2972327080" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1648" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1648"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1107" name="/roberta/encoder/layer.7/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1108" name="/roberta/encoder/layer.7/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1109" name="/roberta/encoder/layer.7/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.7/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1110" name="/roberta/encoder/layer.7/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="1111" name="Constant_1463" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1463" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1112" name="/roberta/encoder/layer.7/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant, /roberta/encoder/layer.7/attention/self/Gather, Constant_1463" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="1113" name="Constant_1469" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1469" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1114" name="/roberta/encoder/layer.7/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Unsqueeze, Constant_1469" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.7/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1115" name="/roberta/encoder/layer.7/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.7/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1116" name="/roberta/encoder/layer.7/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="1117" name="Constant_1467" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1467" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1118" name="/roberta/encoder/layer.7/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_1, /roberta/encoder/layer.7/attention/self/Gather_1, Constant_1467" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="1119" name="Constant_1471" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1471" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1120" name="/roberta/encoder/layer.7/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Unsqueeze_1, Constant_1471" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.7/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1121" name="/roberta/encoder/layer.7/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1122" name="/roberta/encoder/layer.7/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1123" name="/roberta/encoder/layer.7/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Concat, /roberta/encoder/layer.7/attention/self/Constant_2, /roberta/encoder/layer.7/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.7/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1124" name="/roberta/encoder/layer.7/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1125" name="Constant_1545" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1545" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1126" name="/roberta/encoder/layer.7/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1127" name="/roberta/encoder/layer.7/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1128" name="/roberta/encoder/layer.7/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="1129" name="/roberta/encoder/layer.7/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1130" name="/roberta/encoder/layer.7/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1131" name="/roberta/encoder/layer.7/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1132" name="roberta.encoder.layer.7.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2974686376" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1133" name="onnx::MatMul_1651" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2974689448" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1651" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1651"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1134" name="/roberta/encoder/layer.7/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1135" name="/roberta/encoder/layer.7/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1136" name="/roberta/encoder/layer.7/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.7/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1137" name="/roberta/encoder/layer.7/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="1138" name="Constant_1491" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1491" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1139" name="/roberta/encoder/layer.7/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_4, /roberta/encoder/layer.7/attention/self/Gather_2, Constant_1491" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="1140" name="Constant_1497" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1497" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1141" name="/roberta/encoder/layer.7/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Unsqueeze_2, Constant_1497" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.7/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1142" name="/roberta/encoder/layer.7/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.7/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1143" name="/roberta/encoder/layer.7/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="1144" name="Constant_1495" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1495" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1145" name="/roberta/encoder/layer.7/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_5, /roberta/encoder/layer.7/attention/self/Gather_3, Constant_1495" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="1146" name="Constant_1499" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1499" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1147" name="/roberta/encoder/layer.7/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Unsqueeze_3, Constant_1499" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.7/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1148" name="/roberta/encoder/layer.7/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1149" name="/roberta/encoder/layer.7/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1150" name="/roberta/encoder/layer.7/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Concat_1, /roberta/encoder/layer.7/attention/self/Constant_6, /roberta/encoder/layer.7/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.7/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1151" name="/roberta/encoder/layer.7/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1152" name="Constant_1515" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1515" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1153" name="/roberta/encoder/layer.7/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1154" name="/roberta/encoder/layer.7/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1155" name="Constant_1553" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1553" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1156" name="/roberta/encoder/layer.7/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1157" name="/roberta/encoder/layer.7/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.7/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1158" name="/roberta/encoder/layer.7/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="1159" name="Constant_1557" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1557" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1160" name="/roberta/encoder/layer.7/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_13, /roberta/encoder/layer.7/attention/self/Gather_6, Constant_1557" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="1161" name="Constant_1563" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1563" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1162" name="/roberta/encoder/layer.7/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Unsqueeze_6, Constant_1563" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.7/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1163" name="/roberta/encoder/layer.7/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.7/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1164" name="/roberta/encoder/layer.7/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="1165" name="Constant_1561" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1561" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1166" name="/roberta/encoder/layer.7/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_14, /roberta/encoder/layer.7/attention/self/Gather_7, Constant_1561" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="1167" name="Constant_1565" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1565" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1168" name="/roberta/encoder/layer.7/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Unsqueeze_7, Constant_1565" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.7/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1169" name="/roberta/encoder/layer.7/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.7/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1170" name="/roberta/encoder/layer.7/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Concat_3, /roberta/encoder/layer.7/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.7/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1171" name="/roberta/encoder/layer.7/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1172" name="onnx::MatMul_1657" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2977048744" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1657" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1657"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1173" name="/roberta/encoder/layer.7/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1174" name="/roberta/encoder/layer.7/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1175" name="/roberta/encoder/layer.7/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1176" name="Constant_1583" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1583" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1177" name="/roberta/encoder/layer.7/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1178" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1179" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1180" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1181" name="Constant_1588" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1588" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1182" name="/roberta/encoder/layer.7/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1183" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1184" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1185" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1186" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1187" name="roberta.encoder.layer.7.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2979408040" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1188" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1189" name="roberta.encoder.layer.7.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2979411112" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1190" name="/roberta/encoder/layer.7/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1191" name="onnx::MatMul_1658" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="2979414184" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1658" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1658"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1192" name="/roberta/encoder/layer.7/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1193" name="/roberta/encoder/layer.7/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1194" name="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="1195" name="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1196" name="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1197" name="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="1198" name="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1199" name="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1200" name="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="1201" name="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1202" name="onnx::MatMul_1659" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="2988851368" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1659" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1659"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1203" name="/roberta/encoder/layer.7/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1204" name="/roberta/encoder/layer.7/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1205" name="/roberta/encoder/layer.7/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1206" name="Constant_1609" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1609" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1207" name="/roberta/encoder/layer.7/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1208" name="/roberta/encoder/layer.7/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1209" name="/roberta/encoder/layer.7/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1210" name="/roberta/encoder/layer.7/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1211" name="Constant_1614" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1614" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1212" name="/roberta/encoder/layer.7/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1213" name="/roberta/encoder/layer.7/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1214" name="/roberta/encoder/layer.7/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1215" name="/roberta/encoder/layer.7/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1216" name="/roberta/encoder/layer.7/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1217" name="roberta.encoder.layer.7.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2998288552" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1218" name="/roberta/encoder/layer.7/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1219" name="roberta.encoder.layer.7.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="2998291624" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.7.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.7.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1220" name="/roberta/encoder/layer.7/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.7/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.7/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1221" name="onnx::MatMul_1660" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="2998294696" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1660" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1660"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1222" name="/roberta/encoder/layer.8/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1223" name="/roberta/encoder/layer.8/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1224" name="/roberta/encoder/layer.8/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.8/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1225" name="/roberta/encoder/layer.8/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="1226" name="Constant_1684" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1684" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1227" name="/roberta/encoder/layer.8/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_8, /roberta/encoder/layer.8/attention/self/Gather_4, Constant_1684" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="1228" name="Constant_1690" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1690" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1229" name="/roberta/encoder/layer.8/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Unsqueeze_4, Constant_1690" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.8/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1230" name="/roberta/encoder/layer.8/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.8/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1231" name="/roberta/encoder/layer.8/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="1232" name="Constant_1688" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1688" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1233" name="/roberta/encoder/layer.8/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_9, /roberta/encoder/layer.8/attention/self/Gather_5, Constant_1688" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="1234" name="Constant_1692" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1692" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1235" name="/roberta/encoder/layer.8/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Unsqueeze_5, Constant_1692" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.8/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1236" name="/roberta/encoder/layer.8/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1237" name="/roberta/encoder/layer.8/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1238" name="/roberta/encoder/layer.8/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Concat_2, /roberta/encoder/layer.8/attention/self/Constant_10, /roberta/encoder/layer.8/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.8/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1239" name="/roberta/encoder/layer.8/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1240" name="Constant_1708" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1708" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1241" name="/roberta/encoder/layer.8/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1242" name="roberta.encoder.layer.8.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3000653992" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1243" name="onnx::MatMul_1661" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3000657064" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1661" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1661"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1244" name="/roberta/encoder/layer.8/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1245" name="/roberta/encoder/layer.8/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1246" name="/roberta/encoder/layer.8/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.8/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1247" name="/roberta/encoder/layer.8/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="1248" name="Constant_1628" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1628" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1249" name="/roberta/encoder/layer.8/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant, /roberta/encoder/layer.8/attention/self/Gather, Constant_1628" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="1250" name="Constant_1634" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1634" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1251" name="/roberta/encoder/layer.8/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Unsqueeze, Constant_1634" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.8/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1252" name="/roberta/encoder/layer.8/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.8/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1253" name="/roberta/encoder/layer.8/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="1254" name="Constant_1632" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1632" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1255" name="/roberta/encoder/layer.8/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_1, /roberta/encoder/layer.8/attention/self/Gather_1, Constant_1632" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="1256" name="Constant_1636" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1636" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1257" name="/roberta/encoder/layer.8/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Unsqueeze_1, Constant_1636" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.8/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1258" name="/roberta/encoder/layer.8/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1259" name="/roberta/encoder/layer.8/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1260" name="/roberta/encoder/layer.8/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Concat, /roberta/encoder/layer.8/attention/self/Constant_2, /roberta/encoder/layer.8/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.8/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1261" name="/roberta/encoder/layer.8/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1262" name="Constant_1710" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1710" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1263" name="/roberta/encoder/layer.8/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1264" name="/roberta/encoder/layer.8/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1265" name="/roberta/encoder/layer.8/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="1266" name="/roberta/encoder/layer.8/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1267" name="/roberta/encoder/layer.8/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1268" name="/roberta/encoder/layer.8/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1269" name="roberta.encoder.layer.8.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3003016360" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1270" name="onnx::MatMul_1664" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3003019432" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1664" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1664"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1271" name="/roberta/encoder/layer.8/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1272" name="/roberta/encoder/layer.8/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1273" name="/roberta/encoder/layer.8/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.8/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1274" name="/roberta/encoder/layer.8/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="1275" name="Constant_1656" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1656" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1276" name="/roberta/encoder/layer.8/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_4, /roberta/encoder/layer.8/attention/self/Gather_2, Constant_1656" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="1277" name="Constant_1662" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1662" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1278" name="/roberta/encoder/layer.8/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Unsqueeze_2, Constant_1662" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.8/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1279" name="/roberta/encoder/layer.8/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.8/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1280" name="/roberta/encoder/layer.8/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="1281" name="Constant_1660" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1660" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1282" name="/roberta/encoder/layer.8/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_5, /roberta/encoder/layer.8/attention/self/Gather_3, Constant_1660" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="1283" name="Constant_1664" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1664" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1284" name="/roberta/encoder/layer.8/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Unsqueeze_3, Constant_1664" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.8/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1285" name="/roberta/encoder/layer.8/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1286" name="/roberta/encoder/layer.8/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1287" name="/roberta/encoder/layer.8/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Concat_1, /roberta/encoder/layer.8/attention/self/Constant_6, /roberta/encoder/layer.8/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.8/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1288" name="/roberta/encoder/layer.8/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1289" name="Constant_1680" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1680" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1290" name="/roberta/encoder/layer.8/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1291" name="/roberta/encoder/layer.8/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1292" name="Constant_1718" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1718" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1293" name="/roberta/encoder/layer.8/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1294" name="/roberta/encoder/layer.8/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.8/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1295" name="/roberta/encoder/layer.8/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="1296" name="Constant_1722" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1722" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1297" name="/roberta/encoder/layer.8/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_13, /roberta/encoder/layer.8/attention/self/Gather_6, Constant_1722" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="1298" name="Constant_1728" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1728" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1299" name="/roberta/encoder/layer.8/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Unsqueeze_6, Constant_1728" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.8/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1300" name="/roberta/encoder/layer.8/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.8/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1301" name="/roberta/encoder/layer.8/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="1302" name="Constant_1726" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1726" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1303" name="/roberta/encoder/layer.8/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_14, /roberta/encoder/layer.8/attention/self/Gather_7, Constant_1726" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="1304" name="Constant_1730" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1730" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1305" name="/roberta/encoder/layer.8/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Unsqueeze_7, Constant_1730" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.8/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1306" name="/roberta/encoder/layer.8/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.8/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1307" name="/roberta/encoder/layer.8/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Concat_3, /roberta/encoder/layer.8/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.8/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1308" name="/roberta/encoder/layer.8/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1309" name="onnx::MatMul_1670" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3005378728" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1670" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1670"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1310" name="/roberta/encoder/layer.8/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1311" name="/roberta/encoder/layer.8/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1312" name="/roberta/encoder/layer.8/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1313" name="Constant_1748" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1748" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1314" name="/roberta/encoder/layer.8/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1315" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1316" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1317" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1318" name="Constant_1753" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1753" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1319" name="/roberta/encoder/layer.8/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1320" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1321" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1322" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1323" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1324" name="roberta.encoder.layer.8.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3007738024" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1325" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1326" name="roberta.encoder.layer.8.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3007741096" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1327" name="/roberta/encoder/layer.8/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1328" name="onnx::MatMul_1671" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="3007744168" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1671" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1671"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1329" name="/roberta/encoder/layer.8/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1330" name="/roberta/encoder/layer.8/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1331" name="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="1332" name="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1333" name="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1334" name="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="1335" name="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1336" name="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1337" name="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="1338" name="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1339" name="onnx::MatMul_1672" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="3017181352" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1672" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1672"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1340" name="/roberta/encoder/layer.8/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1341" name="/roberta/encoder/layer.8/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1342" name="/roberta/encoder/layer.8/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1343" name="Constant_1774" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1774" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1344" name="/roberta/encoder/layer.8/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1345" name="/roberta/encoder/layer.8/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1346" name="/roberta/encoder/layer.8/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1347" name="/roberta/encoder/layer.8/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1348" name="Constant_1779" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1779" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1349" name="/roberta/encoder/layer.8/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1350" name="/roberta/encoder/layer.8/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1351" name="/roberta/encoder/layer.8/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1352" name="/roberta/encoder/layer.8/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1353" name="/roberta/encoder/layer.8/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1354" name="roberta.encoder.layer.8.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3026618536" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1355" name="/roberta/encoder/layer.8/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1356" name="roberta.encoder.layer.8.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3026621608" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.8.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.8.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1357" name="/roberta/encoder/layer.8/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.8/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.8/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1358" name="onnx::MatMul_1673" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3026624680" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1673" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1673"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1359" name="/roberta/encoder/layer.9/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1360" name="/roberta/encoder/layer.9/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1361" name="/roberta/encoder/layer.9/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.9/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1362" name="/roberta/encoder/layer.9/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="1363" name="Constant_1849" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1849" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1364" name="/roberta/encoder/layer.9/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_8, /roberta/encoder/layer.9/attention/self/Gather_4, Constant_1849" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="1365" name="Constant_1855" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1855" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1366" name="/roberta/encoder/layer.9/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Unsqueeze_4, Constant_1855" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.9/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1367" name="/roberta/encoder/layer.9/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.9/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1368" name="/roberta/encoder/layer.9/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="1369" name="Constant_1853" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1853" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1370" name="/roberta/encoder/layer.9/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_9, /roberta/encoder/layer.9/attention/self/Gather_5, Constant_1853" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="1371" name="Constant_1857" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1857" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1372" name="/roberta/encoder/layer.9/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Unsqueeze_5, Constant_1857" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.9/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1373" name="/roberta/encoder/layer.9/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1374" name="/roberta/encoder/layer.9/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1375" name="/roberta/encoder/layer.9/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Concat_2, /roberta/encoder/layer.9/attention/self/Constant_10, /roberta/encoder/layer.9/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.9/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1376" name="/roberta/encoder/layer.9/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1377" name="Constant_1873" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1873" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1378" name="/roberta/encoder/layer.9/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1379" name="roberta.encoder.layer.9.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3028983976" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1380" name="onnx::MatMul_1674" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3028987048" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1674" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1674"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1381" name="/roberta/encoder/layer.9/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1382" name="/roberta/encoder/layer.9/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1383" name="/roberta/encoder/layer.9/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.9/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1384" name="/roberta/encoder/layer.9/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="1385" name="Constant_1793" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1793" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1386" name="/roberta/encoder/layer.9/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant, /roberta/encoder/layer.9/attention/self/Gather, Constant_1793" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="1387" name="Constant_1799" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1799" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1388" name="/roberta/encoder/layer.9/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Unsqueeze, Constant_1799" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.9/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1389" name="/roberta/encoder/layer.9/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.9/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1390" name="/roberta/encoder/layer.9/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="1391" name="Constant_1797" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1797" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1392" name="/roberta/encoder/layer.9/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_1, /roberta/encoder/layer.9/attention/self/Gather_1, Constant_1797" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="1393" name="Constant_1801" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1801" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1394" name="/roberta/encoder/layer.9/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Unsqueeze_1, Constant_1801" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.9/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1395" name="/roberta/encoder/layer.9/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1396" name="/roberta/encoder/layer.9/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1397" name="/roberta/encoder/layer.9/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Concat, /roberta/encoder/layer.9/attention/self/Constant_2, /roberta/encoder/layer.9/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.9/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1398" name="/roberta/encoder/layer.9/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1399" name="Constant_1875" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1875" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1400" name="/roberta/encoder/layer.9/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1401" name="/roberta/encoder/layer.9/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1402" name="/roberta/encoder/layer.9/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="1403" name="/roberta/encoder/layer.9/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1404" name="/roberta/encoder/layer.9/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1405" name="/roberta/encoder/layer.9/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1406" name="roberta.encoder.layer.9.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3031346344" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1407" name="onnx::MatMul_1677" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3031349416" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1677" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1677"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1408" name="/roberta/encoder/layer.9/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1409" name="/roberta/encoder/layer.9/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1410" name="/roberta/encoder/layer.9/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.9/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1411" name="/roberta/encoder/layer.9/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="1412" name="Constant_1821" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1821" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1413" name="/roberta/encoder/layer.9/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_4, /roberta/encoder/layer.9/attention/self/Gather_2, Constant_1821" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="1414" name="Constant_1827" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1827" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1415" name="/roberta/encoder/layer.9/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Unsqueeze_2, Constant_1827" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.9/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1416" name="/roberta/encoder/layer.9/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.9/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1417" name="/roberta/encoder/layer.9/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="1418" name="Constant_1825" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1825" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1419" name="/roberta/encoder/layer.9/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_5, /roberta/encoder/layer.9/attention/self/Gather_3, Constant_1825" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="1420" name="Constant_1829" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1829" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1421" name="/roberta/encoder/layer.9/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Unsqueeze_3, Constant_1829" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.9/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1422" name="/roberta/encoder/layer.9/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1423" name="/roberta/encoder/layer.9/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1424" name="/roberta/encoder/layer.9/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Concat_1, /roberta/encoder/layer.9/attention/self/Constant_6, /roberta/encoder/layer.9/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.9/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1425" name="/roberta/encoder/layer.9/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1426" name="Constant_1845" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1845" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1427" name="/roberta/encoder/layer.9/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1428" name="/roberta/encoder/layer.9/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1429" name="Constant_1883" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1883" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1430" name="/roberta/encoder/layer.9/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1431" name="/roberta/encoder/layer.9/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.9/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1432" name="/roberta/encoder/layer.9/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="1433" name="Constant_1887" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1887" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1434" name="/roberta/encoder/layer.9/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_13, /roberta/encoder/layer.9/attention/self/Gather_6, Constant_1887" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="1435" name="Constant_1893" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1893" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1436" name="/roberta/encoder/layer.9/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Unsqueeze_6, Constant_1893" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.9/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1437" name="/roberta/encoder/layer.9/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.9/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1438" name="/roberta/encoder/layer.9/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="1439" name="Constant_1891" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1891" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1440" name="/roberta/encoder/layer.9/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_14, /roberta/encoder/layer.9/attention/self/Gather_7, Constant_1891" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="1441" name="Constant_1895" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1895" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1442" name="/roberta/encoder/layer.9/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Unsqueeze_7, Constant_1895" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.9/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1443" name="/roberta/encoder/layer.9/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.9/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1444" name="/roberta/encoder/layer.9/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Concat_3, /roberta/encoder/layer.9/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.9/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1445" name="/roberta/encoder/layer.9/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1446" name="onnx::MatMul_1683" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3033708712" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1683" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1683"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1447" name="/roberta/encoder/layer.9/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1448" name="/roberta/encoder/layer.9/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1449" name="/roberta/encoder/layer.9/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1450" name="Constant_1913" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1913" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1451" name="/roberta/encoder/layer.9/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1452" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1453" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1454" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1455" name="Constant_1918" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1918" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1456" name="/roberta/encoder/layer.9/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1457" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1458" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1459" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1460" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1461" name="roberta.encoder.layer.9.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3036068008" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1462" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1463" name="roberta.encoder.layer.9.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3036071080" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1464" name="/roberta/encoder/layer.9/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1465" name="onnx::MatMul_1684" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="3036074152" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1684" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1684"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1466" name="/roberta/encoder/layer.9/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1467" name="/roberta/encoder/layer.9/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1468" name="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="1469" name="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1470" name="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1471" name="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="1472" name="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1473" name="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1474" name="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="1475" name="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1476" name="onnx::MatMul_1685" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="3045511336" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1685" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1685"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1477" name="/roberta/encoder/layer.9/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1478" name="/roberta/encoder/layer.9/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1479" name="/roberta/encoder/layer.9/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1480" name="Constant_1939" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1939" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1481" name="/roberta/encoder/layer.9/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1482" name="/roberta/encoder/layer.9/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1483" name="/roberta/encoder/layer.9/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1484" name="/roberta/encoder/layer.9/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1485" name="Constant_1944" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1944" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1486" name="/roberta/encoder/layer.9/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1487" name="/roberta/encoder/layer.9/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1488" name="/roberta/encoder/layer.9/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1489" name="/roberta/encoder/layer.9/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1490" name="/roberta/encoder/layer.9/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1491" name="roberta.encoder.layer.9.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3054948520" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1492" name="/roberta/encoder/layer.9/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1493" name="roberta.encoder.layer.9.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3054951592" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.9.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.9.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1494" name="/roberta/encoder/layer.9/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.9/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.9/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1495" name="onnx::MatMul_1686" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3054954664" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1686" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1686"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1496" name="/roberta/encoder/layer.10/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1497" name="/roberta/encoder/layer.10/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1498" name="/roberta/encoder/layer.10/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.10/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1499" name="/roberta/encoder/layer.10/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="1500" name="Constant_2014" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2014" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1501" name="/roberta/encoder/layer.10/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_8, /roberta/encoder/layer.10/attention/self/Gather_4, Constant_2014" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="1502" name="Constant_2020" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2020" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1503" name="/roberta/encoder/layer.10/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Unsqueeze_4, Constant_2020" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.10/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1504" name="/roberta/encoder/layer.10/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.10/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1505" name="/roberta/encoder/layer.10/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="1506" name="Constant_2018" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2018" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1507" name="/roberta/encoder/layer.10/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_9, /roberta/encoder/layer.10/attention/self/Gather_5, Constant_2018" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="1508" name="Constant_2022" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2022" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1509" name="/roberta/encoder/layer.10/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Unsqueeze_5, Constant_2022" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.10/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1510" name="/roberta/encoder/layer.10/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1511" name="/roberta/encoder/layer.10/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1512" name="/roberta/encoder/layer.10/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Concat_2, /roberta/encoder/layer.10/attention/self/Constant_10, /roberta/encoder/layer.10/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.10/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1513" name="/roberta/encoder/layer.10/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1514" name="Constant_2038" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2038" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1515" name="/roberta/encoder/layer.10/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1516" name="roberta.encoder.layer.10.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3057313960" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1517" name="onnx::MatMul_1687" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3057317032" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1687" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1687"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1518" name="/roberta/encoder/layer.10/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1519" name="/roberta/encoder/layer.10/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1520" name="/roberta/encoder/layer.10/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.10/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1521" name="/roberta/encoder/layer.10/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="1522" name="Constant_1958" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1958" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1523" name="/roberta/encoder/layer.10/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant, /roberta/encoder/layer.10/attention/self/Gather, Constant_1958" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="1524" name="Constant_1964" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1964" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1525" name="/roberta/encoder/layer.10/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Unsqueeze, Constant_1964" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.10/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1526" name="/roberta/encoder/layer.10/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.10/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1527" name="/roberta/encoder/layer.10/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="1528" name="Constant_1962" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1962" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1529" name="/roberta/encoder/layer.10/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_1, /roberta/encoder/layer.10/attention/self/Gather_1, Constant_1962" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="1530" name="Constant_1966" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1966" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1531" name="/roberta/encoder/layer.10/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Unsqueeze_1, Constant_1966" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.10/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1532" name="/roberta/encoder/layer.10/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1533" name="/roberta/encoder/layer.10/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1534" name="/roberta/encoder/layer.10/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Concat, /roberta/encoder/layer.10/attention/self/Constant_2, /roberta/encoder/layer.10/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.10/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1535" name="/roberta/encoder/layer.10/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1536" name="Constant_2040" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2040" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1537" name="/roberta/encoder/layer.10/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1538" name="/roberta/encoder/layer.10/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1539" name="/roberta/encoder/layer.10/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="1540" name="/roberta/encoder/layer.10/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1541" name="/roberta/encoder/layer.10/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1542" name="/roberta/encoder/layer.10/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1543" name="roberta.encoder.layer.10.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3059676328" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1544" name="onnx::MatMul_1690" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3059679400" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1690" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1690"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1545" name="/roberta/encoder/layer.10/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1546" name="/roberta/encoder/layer.10/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1547" name="/roberta/encoder/layer.10/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.10/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1548" name="/roberta/encoder/layer.10/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="1549" name="Constant_1986" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1986" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1550" name="/roberta/encoder/layer.10/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_4, /roberta/encoder/layer.10/attention/self/Gather_2, Constant_1986" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="1551" name="Constant_1992" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1992" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1552" name="/roberta/encoder/layer.10/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Unsqueeze_2, Constant_1992" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.10/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1553" name="/roberta/encoder/layer.10/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.10/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1554" name="/roberta/encoder/layer.10/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="1555" name="Constant_1990" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1990" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1556" name="/roberta/encoder/layer.10/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_5, /roberta/encoder/layer.10/attention/self/Gather_3, Constant_1990" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="1557" name="Constant_1994" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_1994" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1558" name="/roberta/encoder/layer.10/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Unsqueeze_3, Constant_1994" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.10/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1559" name="/roberta/encoder/layer.10/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1560" name="/roberta/encoder/layer.10/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1561" name="/roberta/encoder/layer.10/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Concat_1, /roberta/encoder/layer.10/attention/self/Constant_6, /roberta/encoder/layer.10/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.10/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1562" name="/roberta/encoder/layer.10/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1563" name="Constant_2010" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2010" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1564" name="/roberta/encoder/layer.10/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1565" name="/roberta/encoder/layer.10/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1566" name="Constant_2048" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2048" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1567" name="/roberta/encoder/layer.10/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1568" name="/roberta/encoder/layer.10/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.10/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1569" name="/roberta/encoder/layer.10/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="1570" name="Constant_2052" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2052" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1571" name="/roberta/encoder/layer.10/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_13, /roberta/encoder/layer.10/attention/self/Gather_6, Constant_2052" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="1572" name="Constant_2058" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2058" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1573" name="/roberta/encoder/layer.10/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Unsqueeze_6, Constant_2058" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.10/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1574" name="/roberta/encoder/layer.10/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.10/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1575" name="/roberta/encoder/layer.10/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="1576" name="Constant_2056" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2056" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1577" name="/roberta/encoder/layer.10/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_14, /roberta/encoder/layer.10/attention/self/Gather_7, Constant_2056" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="1578" name="Constant_2060" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2060" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1579" name="/roberta/encoder/layer.10/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Unsqueeze_7, Constant_2060" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.10/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1580" name="/roberta/encoder/layer.10/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.10/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1581" name="/roberta/encoder/layer.10/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Concat_3, /roberta/encoder/layer.10/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.10/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1582" name="/roberta/encoder/layer.10/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1583" name="onnx::MatMul_1696" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3062038696" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1696" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1696"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1584" name="/roberta/encoder/layer.10/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1585" name="/roberta/encoder/layer.10/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1586" name="/roberta/encoder/layer.10/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1587" name="Constant_2078" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2078" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1588" name="/roberta/encoder/layer.10/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1589" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1590" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1591" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1592" name="Constant_2083" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2083" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1593" name="/roberta/encoder/layer.10/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1594" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1595" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1596" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1597" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1598" name="roberta.encoder.layer.10.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3064397992" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1599" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1600" name="roberta.encoder.layer.10.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3064401064" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1601" name="/roberta/encoder/layer.10/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1602" name="onnx::MatMul_1697" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="3064404136" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1697" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1697"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1603" name="/roberta/encoder/layer.10/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1604" name="/roberta/encoder/layer.10/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1605" name="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="1606" name="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1607" name="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1608" name="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="1609" name="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1610" name="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1611" name="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="1612" name="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1613" name="onnx::MatMul_1698" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="3073841320" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1698" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1698"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1614" name="/roberta/encoder/layer.10/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1615" name="/roberta/encoder/layer.10/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1616" name="/roberta/encoder/layer.10/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1617" name="Constant_2104" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2104" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1618" name="/roberta/encoder/layer.10/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1619" name="/roberta/encoder/layer.10/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1620" name="/roberta/encoder/layer.10/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1621" name="/roberta/encoder/layer.10/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1622" name="Constant_2109" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2109" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1623" name="/roberta/encoder/layer.10/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1624" name="/roberta/encoder/layer.10/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1625" name="/roberta/encoder/layer.10/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1626" name="/roberta/encoder/layer.10/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1627" name="/roberta/encoder/layer.10/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1628" name="roberta.encoder.layer.10.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3083278504" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1629" name="/roberta/encoder/layer.10/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1630" name="roberta.encoder.layer.10.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3083281576" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.10.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.10.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1631" name="/roberta/encoder/layer.10/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.10/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.10/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1632" name="onnx::MatMul_1699" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3083284648" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1699" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1699"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1633" name="/roberta/encoder/layer.11/attention/self/query/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/query/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/query/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1634" name="/roberta/encoder/layer.11/attention/self/query/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/query/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/query/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1635" name="/roberta/encoder/layer.11/attention/self/Shape_4" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Shape_4" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.11/attention/self/Shape_4_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1636" name="/roberta/encoder/layer.11/attention/self/Constant_8" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_8" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_8_output_0" /> + </output> + </layer> + <layer id="1637" name="Constant_2179" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2179" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1638" name="/roberta/encoder/layer.11/attention/self/Gather_4" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_8, /roberta/encoder/layer.11/attention/self/Gather_4, Constant_2179" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Gather_4_output_0" /> + </output> + </layer> + <layer id="1639" name="Constant_2185" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2185" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1640" name="/roberta/encoder/layer.11/attention/self/Unsqueeze_4" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Unsqueeze_4, Constant_2185" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.11/attention/self/Unsqueeze_4_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1641" name="/roberta/encoder/layer.11/attention/self/Shape_5" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Shape_5" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.11/attention/self/Shape_5_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1642" name="/roberta/encoder/layer.11/attention/self/Constant_9" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_9" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_9_output_0" /> + </output> + </layer> + <layer id="1643" name="Constant_2183" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2183" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1644" name="/roberta/encoder/layer.11/attention/self/Gather_5" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_9, /roberta/encoder/layer.11/attention/self/Gather_5, Constant_2183" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Gather_5_output_0" /> + </output> + </layer> + <layer id="1645" name="Constant_2187" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2187" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1646" name="/roberta/encoder/layer.11/attention/self/Unsqueeze_5" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Unsqueeze_5, Constant_2187" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.11/attention/self/Unsqueeze_5_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1647" name="/roberta/encoder/layer.11/attention/self/Constant_10" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_10" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_10_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1648" name="/roberta/encoder/layer.11/attention/self/Constant_11" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_11" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_11_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1649" name="/roberta/encoder/layer.11/attention/self/Concat_2" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Concat_2, /roberta/encoder/layer.11/attention/self/Constant_10, /roberta/encoder/layer.11/attention/self/Constant_11" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.11/attention/self/Concat_2_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1650" name="/roberta/encoder/layer.11/attention/self/Reshape_2" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Reshape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Reshape_2_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1651" name="Constant_2203" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2203" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1652" name="/roberta/encoder/layer.11/attention/self/Transpose_1" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Transpose_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Transpose_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1653" name="roberta.encoder.layer.11.attention.self.key.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3085643944" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.attention.self.key.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.attention.self.key.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1654" name="onnx::MatMul_1700" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3085647016" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1700" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1700"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1655" name="/roberta/encoder/layer.11/attention/self/key/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/key/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/key/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1656" name="/roberta/encoder/layer.11/attention/self/key/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/key/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/key/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1657" name="/roberta/encoder/layer.11/attention/self/Shape" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Shape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.11/attention/self/Shape_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1658" name="/roberta/encoder/layer.11/attention/self/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_output_0" /> + </output> + </layer> + <layer id="1659" name="Constant_2123" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2123" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1660" name="/roberta/encoder/layer.11/attention/self/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant, /roberta/encoder/layer.11/attention/self/Gather, Constant_2123" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Gather_output_0" /> + </output> + </layer> + <layer id="1661" name="Constant_2129" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2129" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1662" name="/roberta/encoder/layer.11/attention/self/Unsqueeze" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Unsqueeze, Constant_2129" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.11/attention/self/Unsqueeze_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1663" name="/roberta/encoder/layer.11/attention/self/Shape_1" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Shape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.11/attention/self/Shape_1_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1664" name="/roberta/encoder/layer.11/attention/self/Constant_1" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_1_output_0" /> + </output> + </layer> + <layer id="1665" name="Constant_2127" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2127" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1666" name="/roberta/encoder/layer.11/attention/self/Gather_1" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_1, /roberta/encoder/layer.11/attention/self/Gather_1, Constant_2127" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Gather_1_output_0" /> + </output> + </layer> + <layer id="1667" name="Constant_2131" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2131" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1668" name="/roberta/encoder/layer.11/attention/self/Unsqueeze_1" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Unsqueeze_1, Constant_2131" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.11/attention/self/Unsqueeze_1_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1669" name="/roberta/encoder/layer.11/attention/self/Constant_2" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1670" name="/roberta/encoder/layer.11/attention/self/Constant_3" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_3" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1671" name="/roberta/encoder/layer.11/attention/self/Concat" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Concat, /roberta/encoder/layer.11/attention/self/Constant_2, /roberta/encoder/layer.11/attention/self/Constant_3" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.11/attention/self/Concat_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1672" name="/roberta/encoder/layer.11/attention/self/Reshape" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Reshape" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Reshape_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1673" name="Constant_2205" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2776376420" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2205" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1674" name="/roberta/encoder/layer.11/attention/self/Transpose_2" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Transpose_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Transpose_2_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1675" name="/roberta/encoder/layer.11/attention/self/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/MatMul_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1676" name="/roberta/encoder/layer.11/attention/self/Constant_12" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376452" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_12" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Constant_12_output_0" /> + </output> + </layer> + <layer id="1677" name="/roberta/encoder/layer.11/attention/self/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Div_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1678" name="/roberta/encoder/layer.11/attention/self/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>1</dim> + <dim>1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Add_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1679" name="/roberta/encoder/layer.11/attention/self/Softmax" type="SoftMax" version="opset8"> + <data axis="3" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Softmax" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Softmax_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + </output> + </layer> + <layer id="1680" name="roberta.encoder.layer.11.attention.self.value.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3088006312" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.attention.self.value.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.attention.self.value.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1681" name="onnx::MatMul_1703" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3088009384" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1703" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1703"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1682" name="/roberta/encoder/layer.11/attention/self/value/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/value/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/value/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1683" name="/roberta/encoder/layer.11/attention/self/value/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/value/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/value/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1684" name="/roberta/encoder/layer.11/attention/self/Shape_2" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Shape_2" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.11/attention/self/Shape_2_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1685" name="/roberta/encoder/layer.11/attention/self/Constant_4" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_4" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_4_output_0" /> + </output> + </layer> + <layer id="1686" name="Constant_2151" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2151" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1687" name="/roberta/encoder/layer.11/attention/self/Gather_2" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_4, /roberta/encoder/layer.11/attention/self/Gather_2, Constant_2151" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Gather_2_output_0" /> + </output> + </layer> + <layer id="1688" name="Constant_2157" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2157" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1689" name="/roberta/encoder/layer.11/attention/self/Unsqueeze_2" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Unsqueeze_2, Constant_2157" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.11/attention/self/Unsqueeze_2_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1690" name="/roberta/encoder/layer.11/attention/self/Shape_3" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Shape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.11/attention/self/Shape_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1691" name="/roberta/encoder/layer.11/attention/self/Constant_5" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_5" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_5_output_0" /> + </output> + </layer> + <layer id="1692" name="Constant_2155" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2155" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1693" name="/roberta/encoder/layer.11/attention/self/Gather_3" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_5, /roberta/encoder/layer.11/attention/self/Gather_3, Constant_2155" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>3</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Gather_3_output_0" /> + </output> + </layer> + <layer id="1694" name="Constant_2159" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2159" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1695" name="/roberta/encoder/layer.11/attention/self/Unsqueeze_3" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Unsqueeze_3, Constant_2159" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.11/attention/self/Unsqueeze_3_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1696" name="/roberta/encoder/layer.11/attention/self/Constant_6" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014004" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_6" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1697" name="/roberta/encoder/layer.11/attention/self/Constant_7" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2774014012" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_7" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1698" name="/roberta/encoder/layer.11/attention/self/Concat_1" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Concat_1, /roberta/encoder/layer.11/attention/self/Constant_6, /roberta/encoder/layer.11/attention/self/Constant_7" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + <port id="3" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="4" precision="I64" names="/roberta/encoder/layer.11/attention/self/Concat_1_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1699" name="/roberta/encoder/layer.11/attention/self/Reshape_1" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Reshape_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Reshape_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1700" name="Constant_2175" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2175" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1701" name="/roberta/encoder/layer.11/attention/self/Transpose" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Transpose" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Transpose_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1702" name="/roberta/encoder/layer.11/attention/self/MatMul_1" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/MatMul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>-1</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/MatMul_1_output_0"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1703" name="Constant_2213" type="Const" version="opset1"> + <data element_type="i64" shape="4" offset="2774014020" size="32" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2213" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1704" name="/roberta/encoder/layer.11/attention/self/Transpose_3" type="Transpose" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Transpose_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>12</dim> + <dim>-1</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>4</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Transpose_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </output> + </layer> + <layer id="1705" name="/roberta/encoder/layer.11/attention/self/Shape_6" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Shape_6" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.11/attention/self/Shape_6_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1706" name="/roberta/encoder/layer.11/attention/self/Constant_13" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_13" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_13_output_0" /> + </output> + </layer> + <layer id="1707" name="Constant_2217" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2217" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1708" name="/roberta/encoder/layer.11/attention/self/Gather_6" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_13, /roberta/encoder/layer.11/attention/self/Gather_6, Constant_2217" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Gather_6_output_0" /> + </output> + </layer> + <layer id="1709" name="Constant_2223" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2223" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1710" name="/roberta/encoder/layer.11/attention/self/Unsqueeze_6" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Unsqueeze_6, Constant_2223" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.11/attention/self/Unsqueeze_6_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1711" name="/roberta/encoder/layer.11/attention/self/Shape_7" type="ShapeOf" version="opset3"> + <data output_type="i64" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Shape_7" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + </input> + <output> + <port id="1" precision="I64" names="/roberta/encoder/layer.11/attention/self/Shape_7_output_0"> + <dim>4</dim> + </port> + </output> + </layer> + <layer id="1712" name="/roberta/encoder/layer.11/attention/self/Constant_14" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_14" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_14_output_0" /> + </output> + </layer> + <layer id="1713" name="Constant_2221" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2221" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1714" name="/roberta/encoder/layer.11/attention/self/Gather_7" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_14, /roberta/encoder/layer.11/attention/self/Gather_7, Constant_2221" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>4</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Gather_7_output_0" /> + </output> + </layer> + <layer id="1715" name="Constant_2225" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2225" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1716" name="/roberta/encoder/layer.11/attention/self/Unsqueeze_7" type="Unsqueeze" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Unsqueeze_7, Constant_2225" /> + </rt_info> + <input> + <port id="0" precision="I64" /> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="I64" names="/roberta/encoder/layer.11/attention/self/Unsqueeze_7_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1717" name="/roberta/encoder/layer.11/attention/self/Constant_15" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2778738840" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Constant_15" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/encoder/layer.11/attention/self/Constant_15_output_0"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1718" name="/roberta/encoder/layer.11/attention/self/Concat_3" type="Concat" version="opset1"> + <data axis="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Concat_3, /roberta/encoder/layer.11/attention/self/Constant_15" /> + </rt_info> + <input> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + <port id="2" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="3" precision="I64" names="/roberta/encoder/layer.11/attention/self/Concat_3_output_0"> + <dim>3</dim> + </port> + </output> + </layer> + <layer id="1719" name="/roberta/encoder/layer.11/attention/self/Reshape_3" type="Reshape" version="opset1"> + <data special_zero="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/self/Reshape_3" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>12</dim> + <dim>64</dim> + </port> + <port id="1" precision="I64"> + <dim>3</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/self/Reshape_3_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1720" name="onnx::MatMul_1709" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3090368680" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1709" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1709"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1721" name="/roberta/encoder/layer.11/attention/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1722" name="/roberta/encoder/layer.11/attention/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1723" name="/roberta/encoder/layer.11/attention/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1724" name="Constant_2243" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2243" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1725" name="/roberta/encoder/layer.11/attention/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1726" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1727" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1728" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1729" name="Constant_2248" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2248" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1730" name="/roberta/encoder/layer.11/attention/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1731" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1732" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1733" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1734" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1735" name="roberta.encoder.layer.11.attention.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3092727976" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.attention.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.attention.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1736" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1737" name="roberta.encoder.layer.11.attention.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3092731048" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.attention.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.attention.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1738" name="/roberta/encoder/layer.11/attention/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/attention/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/attention/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1739" name="onnx::MatMul_1710" type="Const" version="opset1"> + <data element_type="f32" shape="768, 3072" offset="3092734120" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1710" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1710"> + <dim>768</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1740" name="/roberta/encoder/layer.11/intermediate/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/intermediate/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1741" name="/roberta/encoder/layer.11/intermediate/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/intermediate/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1742" name="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541472" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant_output_0" /> + </output> + </layer> + <layer id="1743" name="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1744" name="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Erf" type="Erf" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Erf" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Erf_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1745" name="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant_1_output_0" /> + </output> + </layer> + <layer id="1746" name="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1747" name="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1748" name="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant_2" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2790541476" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant_2" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Constant_2_output_0" /> + </output> + </layer> + <layer id="1749" name="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Mul_1" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Mul_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/intermediate/intermediate_act_fn/Mul_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + </output> + </layer> + <layer id="1750" name="onnx::MatMul_1711" type="Const" version="opset1"> + <data element_type="f32" shape="3072, 768" offset="3102171304" size="9437184" /> + <rt_info> + <attribute name="fused_names" version="0" value="onnx::MatMul_1711" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="onnx::MatMul_1711"> + <dim>3072</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1751" name="/roberta/encoder/layer.11/output/dense/MatMul" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="false" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/dense/MatMul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>3072</dim> + </port> + <port id="1" precision="FP32"> + <dim>3072</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/dense/MatMul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1752" name="/roberta/encoder/layer.11/output/dense/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/dense/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/dense/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1753" name="/roberta/encoder/layer.11/output/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1754" name="Constant_2269" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2269" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1755" name="/roberta/encoder/layer.11/output/LayerNorm/ReduceMean" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/ReduceMean" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/ReduceMean_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1756" name="/roberta/encoder/layer.11/output/LayerNorm/Sub" type="Subtract" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Sub" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Sub_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1757" name="/roberta/encoder/layer.11/output/LayerNorm/Constant" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648556" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Constant" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Constant_output_0" /> + </output> + </layer> + <layer id="1758" name="/roberta/encoder/layer.11/output/LayerNorm/Pow" type="Power" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Pow" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Pow_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1759" name="Constant_2274" type="Const" version="opset1"> + <data element_type="i64" shape="1" offset="2770069536" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2274" /> + </rt_info> + <output> + <port id="0" precision="I64"> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1760" name="/roberta/encoder/layer.11/output/LayerNorm/ReduceMean_1" type="ReduceMean" version="opset1"> + <data keep_dims="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/ReduceMean_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64"> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/ReduceMean_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1761" name="/roberta/encoder/layer.11/output/LayerNorm/Constant_1" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2771648560" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Constant_1" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Constant_1_output_0" /> + </output> + </layer> + <layer id="1762" name="/roberta/encoder/layer.11/output/LayerNorm/Add" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Add" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Add_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1763" name="/roberta/encoder/layer.11/output/LayerNorm/Sqrt" type="Sqrt" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Sqrt" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Sqrt_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </output> + </layer> + <layer id="1764" name="/roberta/encoder/layer.11/output/LayerNorm/Div" type="Divide" version="opset1"> + <data auto_broadcast="numpy" m_pythondiv="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Div" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>1</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Div_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1765" name="roberta.encoder.layer.11.output.LayerNorm.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3111608488" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.output.LayerNorm.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.output.LayerNorm.weight"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1766" name="/roberta/encoder/layer.11/output/LayerNorm/Mul" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Mul" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Mul_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1767" name="roberta.encoder.layer.11.output.LayerNorm.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3111611560" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="roberta.encoder.layer.11.output.LayerNorm.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="roberta.encoder.layer.11.output.LayerNorm.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1768" name="/roberta/encoder/layer.11/output/LayerNorm/Add_1" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/encoder/layer.11/output/LayerNorm/Add_1" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/roberta/encoder/layer.11/output/LayerNorm/Add_1_output_0"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1769" name="/roberta/Constant" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770062336" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="/roberta/Constant" /> + </rt_info> + <output> + <port id="0" precision="I64" names="/roberta/Constant_output_0" /> + </output> + </layer> + <layer id="1770" name="Constant_2282" type="Const" version="opset1"> + <data element_type="i64" shape="" offset="2770069528" size="8" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2282" /> + </rt_info> + <output> + <port id="0" precision="I64" /> + </output> + </layer> + <layer id="1771" name="/classifier/Gather" type="Gather" version="opset8"> + <data batch_dims="0" /> + <rt_info> + <attribute name="fused_names" version="0" value="/classifier/Gather" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="I64" /> + <port id="2" precision="I64" /> + </input> + <output> + <port id="3" precision="FP32" names="/classifier/Gather_output_0"> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1772" name="classifier.dense.weight" type="Const" version="opset1"> + <data element_type="f32" shape="768, 768" offset="3111614632" size="2359296" /> + <rt_info> + <attribute name="fused_names" version="0" value="classifier.dense.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="classifier.dense.weight"> + <dim>768</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1773" name="/classifier/dense/Gemm/WithoutBiases" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/classifier/dense/Gemm/WithoutBiases" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32"> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1774" name="Constant_2284" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2284" /> + </rt_info> + <output> + <port id="0" precision="FP32" /> + </output> + </layer> + <layer id="1775" name="Multiply_2287" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="Multiply_2287" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32"> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1776" name="Constant_2285" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2285" /> + </rt_info> + <output> + <port id="0" precision="FP32" /> + </output> + </layer> + <layer id="1777" name="classifier.dense.bias" type="Const" version="opset1"> + <data element_type="f32" shape="768" offset="3113973928" size="3072" /> + <rt_info> + <attribute name="fused_names" version="0" value="classifier.dense.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="classifier.dense.bias"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1778" name="Multiply_2288" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="Multiply_2288" /> + </rt_info> + <input> + <port id="0" precision="FP32" /> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32"> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1779" name="/classifier/dense/Gemm" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="/classifier/dense/Gemm" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="/classifier/dense/Gemm_output_0"> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1780" name="/classifier/Tanh" type="Tanh" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="/classifier/Tanh" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="1" precision="FP32" names="/classifier/Tanh_output_0"> + <dim>-1</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1781" name="classifier.out_proj.weight" type="Const" version="opset1"> + <data element_type="f32" shape="102, 768" offset="3113977000" size="313344" /> + <rt_info> + <attribute name="fused_names" version="0" value="classifier.out_proj.weight" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="classifier.out_proj.weight"> + <dim>102</dim> + <dim>768</dim> + </port> + </output> + </layer> + <layer id="1782" name="/classifier/out_proj/Gemm/WithoutBiases" type="MatMul" version="opset1"> + <data transpose_a="false" transpose_b="true" /> + <rt_info> + <attribute name="fused_names" version="0" value="/classifier/out_proj/Gemm/WithoutBiases" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>768</dim> + </port> + <port id="1" precision="FP32"> + <dim>102</dim> + <dim>768</dim> + </port> + </input> + <output> + <port id="2" precision="FP32"> + <dim>-1</dim> + <dim>102</dim> + </port> + </output> + </layer> + <layer id="1783" name="Constant_2291" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2291" /> + </rt_info> + <output> + <port id="0" precision="FP32" /> + </output> + </layer> + <layer id="1784" name="Multiply_2294" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="Multiply_2294" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>102</dim> + </port> + <port id="1" precision="FP32" /> + </input> + <output> + <port id="2" precision="FP32"> + <dim>-1</dim> + <dim>102</dim> + </port> + </output> + </layer> + <layer id="1785" name="Constant_2292" type="Const" version="opset1"> + <data element_type="f32" shape="" offset="2776376456" size="4" /> + <rt_info> + <attribute name="fused_names" version="0" value="Constant_2292" /> + </rt_info> + <output> + <port id="0" precision="FP32" /> + </output> + </layer> + <layer id="1786" name="classifier.out_proj.bias" type="Const" version="opset1"> + <data element_type="f32" shape="102" offset="3114290344" size="408" /> + <rt_info> + <attribute name="fused_names" version="0" value="classifier.out_proj.bias" /> + </rt_info> + <output> + <port id="0" precision="FP32" names="classifier.out_proj.bias"> + <dim>102</dim> + </port> + </output> + </layer> + <layer id="1787" name="Multiply_2295" type="Multiply" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="Multiply_2295" /> + </rt_info> + <input> + <port id="0" precision="FP32" /> + <port id="1" precision="FP32"> + <dim>102</dim> + </port> + </input> + <output> + <port id="2" precision="FP32"> + <dim>102</dim> + </port> + </output> + </layer> + <layer id="1788" name="logits" type="Add" version="opset1"> + <data auto_broadcast="numpy" /> + <rt_info> + <attribute name="fused_names" version="0" value="logits" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>102</dim> + </port> + <port id="1" precision="FP32"> + <dim>102</dim> + </port> + </input> + <output> + <port id="2" precision="FP32" names="logits"> + <dim>-1</dim> + <dim>102</dim> + </port> + </output> + </layer> + <layer id="1789" name="logits/sink_port_0" type="Result" version="opset1"> + <rt_info> + <attribute name="fused_names" version="0" value="logits/sink_port_0" /> + </rt_info> + <input> + <port id="0" precision="FP32"> + <dim>-1</dim> + <dim>102</dim> + </port> + </input> + </layer> + </layers> + <edges> + <edge from-layer="0" from-port="0" to-layer="164" to-port="0" /> + <edge from-layer="1" from-port="0" to-layer="65" to-port="0" /> + <edge from-layer="1" from-port="0" to-layer="89" to-port="0" /> + <edge from-layer="1" from-port="0" to-layer="52" to-port="1" /> + <edge from-layer="1" from-port="0" to-layer="56" to-port="0" /> + <edge from-layer="2" from-port="0" to-layer="1752" to-port="0" /> + <edge from-layer="3" from-port="0" to-layer="1741" to-port="0" /> + <edge from-layer="4" from-port="0" to-layer="1722" to-port="0" /> + <edge from-layer="5" from-port="0" to-layer="1634" to-port="0" /> + <edge from-layer="6" from-port="0" to-layer="1615" to-port="0" /> + <edge from-layer="7" from-port="0" to-layer="1604" to-port="0" /> + <edge from-layer="8" from-port="0" to-layer="1585" to-port="0" /> + <edge from-layer="9" from-port="0" to-layer="1497" to-port="0" /> + <edge from-layer="10" from-port="0" to-layer="1478" to-port="0" /> + <edge from-layer="11" from-port="0" to-layer="1467" to-port="0" /> + <edge from-layer="12" from-port="0" to-layer="1448" to-port="0" /> + <edge from-layer="13" from-port="0" to-layer="1360" to-port="0" /> + <edge from-layer="14" from-port="0" to-layer="1341" to-port="0" /> + <edge from-layer="15" from-port="0" to-layer="1330" to-port="0" /> + <edge from-layer="16" from-port="0" to-layer="1311" to-port="0" /> + <edge from-layer="17" from-port="0" to-layer="1223" to-port="0" /> + <edge from-layer="18" from-port="0" to-layer="1204" to-port="0" /> + <edge from-layer="19" from-port="0" to-layer="1193" to-port="0" /> + <edge from-layer="20" from-port="0" to-layer="1174" to-port="0" /> + <edge from-layer="21" from-port="0" to-layer="1086" to-port="0" /> + <edge from-layer="22" from-port="0" to-layer="1067" to-port="0" /> + <edge from-layer="23" from-port="0" to-layer="1056" to-port="0" /> + <edge from-layer="24" from-port="0" to-layer="1037" to-port="0" /> + <edge from-layer="25" from-port="0" to-layer="949" to-port="0" /> + <edge from-layer="26" from-port="0" to-layer="930" to-port="0" /> + <edge from-layer="27" from-port="0" to-layer="919" to-port="0" /> + <edge from-layer="28" from-port="0" to-layer="900" to-port="0" /> + <edge from-layer="29" from-port="0" to-layer="812" to-port="0" /> + <edge from-layer="30" from-port="0" to-layer="793" to-port="0" /> + <edge from-layer="31" from-port="0" to-layer="782" to-port="0" /> + <edge from-layer="32" from-port="0" to-layer="763" to-port="0" /> + <edge from-layer="33" from-port="0" to-layer="675" to-port="0" /> + <edge from-layer="34" from-port="0" to-layer="656" to-port="0" /> + <edge from-layer="35" from-port="0" to-layer="645" to-port="0" /> + <edge from-layer="36" from-port="0" to-layer="626" to-port="0" /> + <edge from-layer="37" from-port="0" to-layer="538" to-port="0" /> + <edge from-layer="38" from-port="0" to-layer="519" to-port="0" /> + <edge from-layer="39" from-port="0" to-layer="508" to-port="0" /> + <edge from-layer="40" from-port="0" to-layer="489" to-port="0" /> + <edge from-layer="41" from-port="0" to-layer="401" to-port="0" /> + <edge from-layer="42" from-port="0" to-layer="382" to-port="0" /> + <edge from-layer="43" from-port="0" to-layer="371" to-port="0" /> + <edge from-layer="44" from-port="0" to-layer="352" to-port="0" /> + <edge from-layer="45" from-port="0" to-layer="264" to-port="0" /> + <edge from-layer="46" from-port="0" to-layer="245" to-port="0" /> + <edge from-layer="47" from-port="0" to-layer="234" to-port="0" /> + <edge from-layer="48" from-port="0" to-layer="215" to-port="0" /> + <edge from-layer="49" from-port="0" to-layer="118" to-port="0" /> + <edge from-layer="50" from-port="0" to-layer="52" to-port="0" /> + <edge from-layer="51" from-port="0" to-layer="52" to-port="2" /> + <edge from-layer="52" from-port="3" to-layer="86" to-port="0" /> + <edge from-layer="53" from-port="0" to-layer="85" to-port="0" /> + <edge from-layer="54" from-port="0" to-layer="64" to-port="0" /> + <edge from-layer="55" from-port="0" to-layer="64" to-port="1" /> + <edge from-layer="56" from-port="1" to-layer="59" to-port="0" /> + <edge from-layer="57" from-port="0" to-layer="59" to-port="1" /> + <edge from-layer="58" from-port="0" to-layer="59" to-port="2" /> + <edge from-layer="59" from-port="3" to-layer="61" to-port="0" /> + <edge from-layer="59" from-port="3" to-layer="72" to-port="0" /> + <edge from-layer="60" from-port="0" to-layer="61" to-port="1" /> + <edge from-layer="61" from-port="2" to-layer="64" to-port="2" /> + <edge from-layer="62" from-port="0" to-layer="64" to-port="3" /> + <edge from-layer="63" from-port="0" to-layer="64" to-port="4" /> + <edge from-layer="64" from-port="5" to-layer="83" to-port="0" /> + <edge from-layer="65" from-port="1" to-layer="68" to-port="0" /> + <edge from-layer="66" from-port="0" to-layer="68" to-port="1" /> + <edge from-layer="67" from-port="0" to-layer="68" to-port="2" /> + <edge from-layer="68" from-port="3" to-layer="70" to-port="0" /> + <edge from-layer="69" from-port="0" to-layer="70" to-port="1" /> + <edge from-layer="70" from-port="2" to-layer="73" to-port="0" /> + <edge from-layer="71" from-port="0" to-layer="72" to-port="1" /> + <edge from-layer="72" from-port="2" to-layer="73" to-port="1" /> + <edge from-layer="73" from-port="2" to-layer="75" to-port="0" /> + <edge from-layer="74" from-port="0" to-layer="75" to-port="1" /> + <edge from-layer="75" from-port="2" to-layer="77" to-port="0" /> + <edge from-layer="75" from-port="2" to-layer="82" to-port="2" /> + <edge from-layer="75" from-port="2" to-layer="81" to-port="0" /> + <edge from-layer="76" from-port="0" to-layer="78" to-port="0" /> + <edge from-layer="77" from-port="1" to-layer="78" to-port="1" /> + <edge from-layer="78" from-port="2" to-layer="80" to-port="0" /> + <edge from-layer="78" from-port="2" to-layer="82" to-port="1" /> + <edge from-layer="79" from-port="0" to-layer="80" to-port="1" /> + <edge from-layer="80" from-port="2" to-layer="81" to-port="1" /> + <edge from-layer="81" from-port="2" to-layer="82" to-port="0" /> + <edge from-layer="82" from-port="3" to-layer="83" to-port="1" /> + <edge from-layer="83" from-port="2" to-layer="85" to-port="1" /> + <edge from-layer="84" from-port="0" to-layer="85" to-port="2" /> + <edge from-layer="85" from-port="3" to-layer="86" to-port="1" /> + <edge from-layer="86" from-port="2" to-layer="100" to-port="0" /> + <edge from-layer="87" from-port="0" to-layer="99" to-port="0" /> + <edge from-layer="88" from-port="0" to-layer="89" to-port="1" /> + <edge from-layer="89" from-port="2" to-layer="90" to-port="0" /> + <edge from-layer="90" from-port="1" to-layer="91" to-port="0" /> + <edge from-layer="91" from-port="1" to-layer="93" to-port="0" /> + <edge from-layer="91" from-port="1" to-layer="94" to-port="1" /> + <edge from-layer="92" from-port="0" to-layer="93" to-port="1" /> + <edge from-layer="93" from-port="2" to-layer="94" to-port="0" /> + <edge from-layer="94" from-port="2" to-layer="95" to-port="0" /> + <edge from-layer="95" from-port="1" to-layer="97" to-port="0" /> + <edge from-layer="96" from-port="0" to-layer="97" to-port="1" /> + <edge from-layer="97" from-port="2" to-layer="99" to-port="1" /> + <edge from-layer="98" from-port="0" to-layer="99" to-port="2" /> + <edge from-layer="99" from-port="3" to-layer="100" to-port="1" /> + <edge from-layer="100" from-port="2" to-layer="102" to-port="0" /> + <edge from-layer="100" from-port="2" to-layer="103" to-port="0" /> + <edge from-layer="101" from-port="0" to-layer="102" to-port="1" /> + <edge from-layer="102" from-port="2" to-layer="103" to-port="1" /> + <edge from-layer="103" from-port="2" to-layer="111" to-port="0" /> + <edge from-layer="103" from-port="2" to-layer="105" to-port="0" /> + <edge from-layer="104" from-port="0" to-layer="105" to-port="1" /> + <edge from-layer="105" from-port="2" to-layer="107" to-port="0" /> + <edge from-layer="106" from-port="0" to-layer="107" to-port="1" /> + <edge from-layer="107" from-port="2" to-layer="109" to-port="0" /> + <edge from-layer="108" from-port="0" to-layer="109" to-port="1" /> + <edge from-layer="109" from-port="2" to-layer="110" to-port="0" /> + <edge from-layer="110" from-port="1" to-layer="111" to-port="1" /> + <edge from-layer="111" from-port="2" to-layer="113" to-port="0" /> + <edge from-layer="112" from-port="0" to-layer="113" to-port="1" /> + <edge from-layer="113" from-port="2" to-layer="115" to-port="0" /> + <edge from-layer="114" from-port="0" to-layer="115" to-port="1" /> + <edge from-layer="115" from-port="2" to-layer="117" to-port="0" /> + <edge from-layer="115" from-port="2" to-layer="175" to-port="0" /> + <edge from-layer="115" from-port="2" to-layer="139" to-port="0" /> + <edge from-layer="115" from-port="2" to-layer="216" to-port="1" /> + <edge from-layer="116" from-port="0" to-layer="117" to-port="1" /> + <edge from-layer="117" from-port="2" to-layer="118" to-port="1" /> + <edge from-layer="118" from-port="2" to-layer="125" to-port="0" /> + <edge from-layer="118" from-port="2" to-layer="134" to-port="0" /> + <edge from-layer="118" from-port="2" to-layer="119" to-port="0" /> + <edge from-layer="119" from-port="1" to-layer="122" to-port="0" /> + <edge from-layer="120" from-port="0" to-layer="122" to-port="1" /> + <edge from-layer="121" from-port="0" to-layer="122" to-port="2" /> + <edge from-layer="122" from-port="3" to-layer="124" to-port="0" /> + <edge from-layer="123" from-port="0" to-layer="124" to-port="1" /> + <edge from-layer="124" from-port="2" to-layer="133" to-port="0" /> + <edge from-layer="125" from-port="1" to-layer="128" to-port="0" /> + <edge from-layer="126" from-port="0" to-layer="128" to-port="1" /> + <edge from-layer="127" from-port="0" to-layer="128" to-port="2" /> + <edge from-layer="128" from-port="3" to-layer="130" to-port="0" /> + <edge from-layer="129" from-port="0" to-layer="130" to-port="1" /> + <edge from-layer="130" from-port="2" to-layer="133" to-port="1" /> + <edge from-layer="131" from-port="0" to-layer="133" to-port="2" /> + <edge from-layer="132" from-port="0" to-layer="133" to-port="3" /> + <edge from-layer="133" from-port="4" to-layer="134" to-port="1" /> + <edge from-layer="134" from-port="2" to-layer="136" to-port="0" /> + <edge from-layer="135" from-port="0" to-layer="136" to-port="1" /> + <edge from-layer="136" from-port="2" to-layer="159" to-port="0" /> + <edge from-layer="137" from-port="0" to-layer="140" to-port="0" /> + <edge from-layer="138" from-port="0" to-layer="139" to-port="1" /> + <edge from-layer="139" from-port="2" to-layer="140" to-port="1" /> + <edge from-layer="140" from-port="2" to-layer="141" to-port="0" /> + <edge from-layer="140" from-port="2" to-layer="147" to-port="0" /> + <edge from-layer="140" from-port="2" to-layer="156" to-port="0" /> + <edge from-layer="141" from-port="1" to-layer="144" to-port="0" /> + <edge from-layer="142" from-port="0" to-layer="144" to-port="1" /> + <edge from-layer="143" from-port="0" to-layer="144" to-port="2" /> + <edge from-layer="144" from-port="3" to-layer="146" to-port="0" /> + <edge from-layer="145" from-port="0" to-layer="146" to-port="1" /> + <edge from-layer="146" from-port="2" to-layer="155" to-port="0" /> + <edge from-layer="147" from-port="1" to-layer="150" to-port="0" /> + <edge from-layer="148" from-port="0" to-layer="150" to-port="1" /> + <edge from-layer="149" from-port="0" to-layer="150" to-port="2" /> + <edge from-layer="150" from-port="3" to-layer="152" to-port="0" /> + <edge from-layer="151" from-port="0" to-layer="152" to-port="1" /> + <edge from-layer="152" from-port="2" to-layer="155" to-port="1" /> + <edge from-layer="153" from-port="0" to-layer="155" to-port="2" /> + <edge from-layer="154" from-port="0" to-layer="155" to-port="3" /> + <edge from-layer="155" from-port="4" to-layer="156" to-port="1" /> + <edge from-layer="156" from-port="2" to-layer="158" to-port="0" /> + <edge from-layer="157" from-port="0" to-layer="158" to-port="1" /> + <edge from-layer="158" from-port="2" to-layer="159" to-port="1" /> + <edge from-layer="159" from-port="2" to-layer="161" to-port="0" /> + <edge from-layer="160" from-port="0" to-layer="161" to-port="1" /> + <edge from-layer="161" from-port="2" to-layer="171" to-port="0" /> + <edge from-layer="162" from-port="0" to-layer="168" to-port="0" /> + <edge from-layer="163" from-port="0" to-layer="164" to-port="1" /> + <edge from-layer="164" from-port="2" to-layer="166" to-port="0" /> + <edge from-layer="165" from-port="0" to-layer="166" to-port="1" /> + <edge from-layer="166" from-port="2" to-layer="167" to-port="0" /> + <edge from-layer="167" from-port="1" to-layer="168" to-port="1" /> + <edge from-layer="168" from-port="2" to-layer="170" to-port="0" /> + <edge from-layer="169" from-port="0" to-layer="170" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="308" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="582" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="445" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="993" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="1130" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="1267" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="1404" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="1541" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="1678" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="856" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="719" to-port="1" /> + <edge from-layer="170" from-port="2" to-layer="171" to-port="1" /> + <edge from-layer="171" from-port="2" to-layer="172" to-port="0" /> + <edge from-layer="172" from-port="1" to-layer="195" to-port="0" /> + <edge from-layer="173" from-port="0" to-layer="176" to-port="0" /> + <edge from-layer="174" from-port="0" to-layer="175" to-port="1" /> + <edge from-layer="175" from-port="2" to-layer="176" to-port="1" /> + <edge from-layer="176" from-port="2" to-layer="177" to-port="0" /> + <edge from-layer="176" from-port="2" to-layer="183" to-port="0" /> + <edge from-layer="176" from-port="2" to-layer="192" to-port="0" /> + <edge from-layer="177" from-port="1" to-layer="180" to-port="0" /> + <edge from-layer="178" from-port="0" to-layer="180" to-port="1" /> + <edge from-layer="179" from-port="0" to-layer="180" to-port="2" /> + <edge from-layer="180" from-port="3" to-layer="182" to-port="0" /> + <edge from-layer="181" from-port="0" to-layer="182" to-port="1" /> + <edge from-layer="182" from-port="2" to-layer="191" to-port="0" /> + <edge from-layer="183" from-port="1" to-layer="186" to-port="0" /> + <edge from-layer="184" from-port="0" to-layer="186" to-port="1" /> + <edge from-layer="185" from-port="0" to-layer="186" to-port="2" /> + <edge from-layer="186" from-port="3" to-layer="188" to-port="0" /> + <edge from-layer="187" from-port="0" to-layer="188" to-port="1" /> + <edge from-layer="188" from-port="2" to-layer="191" to-port="1" /> + <edge from-layer="189" from-port="0" to-layer="191" to-port="2" /> + <edge from-layer="190" from-port="0" to-layer="191" to-port="3" /> + <edge from-layer="191" from-port="4" to-layer="192" to-port="1" /> + <edge from-layer="192" from-port="2" to-layer="194" to-port="0" /> + <edge from-layer="193" from-port="0" to-layer="194" to-port="1" /> + <edge from-layer="194" from-port="2" to-layer="195" to-port="1" /> + <edge from-layer="195" from-port="2" to-layer="197" to-port="0" /> + <edge from-layer="196" from-port="0" to-layer="197" to-port="1" /> + <edge from-layer="197" from-port="2" to-layer="212" to-port="0" /> + <edge from-layer="197" from-port="2" to-layer="198" to-port="0" /> + <edge from-layer="197" from-port="2" to-layer="204" to-port="0" /> + <edge from-layer="198" from-port="1" to-layer="201" to-port="0" /> + <edge from-layer="199" from-port="0" to-layer="201" to-port="1" /> + <edge from-layer="200" from-port="0" to-layer="201" to-port="2" /> + <edge from-layer="201" from-port="3" to-layer="203" to-port="0" /> + <edge from-layer="202" from-port="0" to-layer="203" to-port="1" /> + <edge from-layer="203" from-port="2" to-layer="211" to-port="0" /> + <edge from-layer="204" from-port="1" to-layer="207" to-port="0" /> + <edge from-layer="205" from-port="0" to-layer="207" to-port="1" /> + <edge from-layer="206" from-port="0" to-layer="207" to-port="2" /> + <edge from-layer="207" from-port="3" to-layer="209" to-port="0" /> + <edge from-layer="208" from-port="0" to-layer="209" to-port="1" /> + <edge from-layer="209" from-port="2" to-layer="211" to-port="1" /> + <edge from-layer="210" from-port="0" to-layer="211" to-port="2" /> + <edge from-layer="211" from-port="3" to-layer="212" to-port="1" /> + <edge from-layer="212" from-port="2" to-layer="214" to-port="0" /> + <edge from-layer="213" from-port="0" to-layer="214" to-port="1" /> + <edge from-layer="214" from-port="2" to-layer="215" to-port="1" /> + <edge from-layer="215" from-port="2" to-layer="216" to-port="0" /> + <edge from-layer="216" from-port="2" to-layer="218" to-port="0" /> + <edge from-layer="216" from-port="2" to-layer="219" to-port="0" /> + <edge from-layer="217" from-port="0" to-layer="218" to-port="1" /> + <edge from-layer="218" from-port="2" to-layer="219" to-port="1" /> + <edge from-layer="219" from-port="2" to-layer="221" to-port="0" /> + <edge from-layer="219" from-port="2" to-layer="227" to-port="0" /> + <edge from-layer="220" from-port="0" to-layer="221" to-port="1" /> + <edge from-layer="221" from-port="2" to-layer="223" to-port="0" /> + <edge from-layer="222" from-port="0" to-layer="223" to-port="1" /> + <edge from-layer="223" from-port="2" to-layer="225" to-port="0" /> + <edge from-layer="224" from-port="0" to-layer="225" to-port="1" /> + <edge from-layer="225" from-port="2" to-layer="226" to-port="0" /> + <edge from-layer="226" from-port="1" to-layer="227" to-port="1" /> + <edge from-layer="227" from-port="2" to-layer="229" to-port="0" /> + <edge from-layer="228" from-port="0" to-layer="229" to-port="1" /> + <edge from-layer="229" from-port="2" to-layer="231" to-port="0" /> + <edge from-layer="230" from-port="0" to-layer="231" to-port="1" /> + <edge from-layer="231" from-port="2" to-layer="246" to-port="1" /> + <edge from-layer="231" from-port="2" to-layer="233" to-port="0" /> + <edge from-layer="232" from-port="0" to-layer="233" to-port="1" /> + <edge from-layer="233" from-port="2" to-layer="234" to-port="1" /> + <edge from-layer="234" from-port="2" to-layer="236" to-port="0" /> + <edge from-layer="234" from-port="2" to-layer="240" to-port="0" /> + <edge from-layer="235" from-port="0" to-layer="236" to-port="1" /> + <edge from-layer="236" from-port="2" to-layer="237" to-port="0" /> + <edge from-layer="237" from-port="1" to-layer="239" to-port="0" /> + <edge from-layer="238" from-port="0" to-layer="239" to-port="1" /> + <edge from-layer="239" from-port="2" to-layer="240" to-port="1" /> + <edge from-layer="240" from-port="2" to-layer="242" to-port="0" /> + <edge from-layer="241" from-port="0" to-layer="242" to-port="1" /> + <edge from-layer="242" from-port="2" to-layer="244" to-port="0" /> + <edge from-layer="243" from-port="0" to-layer="244" to-port="1" /> + <edge from-layer="244" from-port="2" to-layer="245" to-port="1" /> + <edge from-layer="245" from-port="2" to-layer="246" to-port="0" /> + <edge from-layer="246" from-port="2" to-layer="248" to-port="0" /> + <edge from-layer="246" from-port="2" to-layer="249" to-port="0" /> + <edge from-layer="247" from-port="0" to-layer="248" to-port="1" /> + <edge from-layer="248" from-port="2" to-layer="249" to-port="1" /> + <edge from-layer="249" from-port="2" to-layer="251" to-port="0" /> + <edge from-layer="249" from-port="2" to-layer="257" to-port="0" /> + <edge from-layer="250" from-port="0" to-layer="251" to-port="1" /> + <edge from-layer="251" from-port="2" to-layer="253" to-port="0" /> + <edge from-layer="252" from-port="0" to-layer="253" to-port="1" /> + <edge from-layer="253" from-port="2" to-layer="255" to-port="0" /> + <edge from-layer="254" from-port="0" to-layer="255" to-port="1" /> + <edge from-layer="255" from-port="2" to-layer="256" to-port="0" /> + <edge from-layer="256" from-port="1" to-layer="257" to-port="1" /> + <edge from-layer="257" from-port="2" to-layer="259" to-port="0" /> + <edge from-layer="258" from-port="0" to-layer="259" to-port="1" /> + <edge from-layer="259" from-port="2" to-layer="261" to-port="0" /> + <edge from-layer="260" from-port="0" to-layer="261" to-port="1" /> + <edge from-layer="261" from-port="2" to-layer="263" to-port="0" /> + <edge from-layer="261" from-port="2" to-layer="353" to-port="1" /> + <edge from-layer="261" from-port="2" to-layer="312" to-port="0" /> + <edge from-layer="261" from-port="2" to-layer="285" to-port="0" /> + <edge from-layer="262" from-port="0" to-layer="263" to-port="1" /> + <edge from-layer="263" from-port="2" to-layer="264" to-port="1" /> + <edge from-layer="264" from-port="2" to-layer="280" to-port="0" /> + <edge from-layer="264" from-port="2" to-layer="265" to-port="0" /> + <edge from-layer="264" from-port="2" to-layer="271" to-port="0" /> + <edge from-layer="265" from-port="1" to-layer="268" to-port="0" /> + <edge from-layer="266" from-port="0" to-layer="268" to-port="1" /> + <edge from-layer="267" from-port="0" to-layer="268" to-port="2" /> + <edge from-layer="268" from-port="3" to-layer="270" to-port="0" /> + <edge from-layer="269" from-port="0" to-layer="270" to-port="1" /> + <edge from-layer="270" from-port="2" to-layer="279" to-port="0" /> + <edge from-layer="271" from-port="1" to-layer="274" to-port="0" /> + <edge from-layer="272" from-port="0" to-layer="274" to-port="1" /> + <edge from-layer="273" from-port="0" to-layer="274" to-port="2" /> + <edge from-layer="274" from-port="3" to-layer="276" to-port="0" /> + <edge from-layer="275" from-port="0" to-layer="276" to-port="1" /> + <edge from-layer="276" from-port="2" to-layer="279" to-port="1" /> + <edge from-layer="277" from-port="0" to-layer="279" to-port="2" /> + <edge from-layer="278" from-port="0" to-layer="279" to-port="3" /> + <edge from-layer="279" from-port="4" to-layer="280" to-port="1" /> + <edge from-layer="280" from-port="2" to-layer="282" to-port="0" /> + <edge from-layer="281" from-port="0" to-layer="282" to-port="1" /> + <edge from-layer="282" from-port="2" to-layer="305" to-port="0" /> + <edge from-layer="283" from-port="0" to-layer="286" to-port="0" /> + <edge from-layer="284" from-port="0" to-layer="285" to-port="1" /> + <edge from-layer="285" from-port="2" to-layer="286" to-port="1" /> + <edge from-layer="286" from-port="2" to-layer="293" to-port="0" /> + <edge from-layer="286" from-port="2" to-layer="287" to-port="0" /> + <edge from-layer="286" from-port="2" to-layer="302" to-port="0" /> + <edge from-layer="287" from-port="1" to-layer="290" to-port="0" /> + <edge from-layer="288" from-port="0" to-layer="290" to-port="1" /> + <edge from-layer="289" from-port="0" to-layer="290" to-port="2" /> + <edge from-layer="290" from-port="3" to-layer="292" to-port="0" /> + <edge from-layer="291" from-port="0" to-layer="292" to-port="1" /> + <edge from-layer="292" from-port="2" to-layer="301" to-port="0" /> + <edge from-layer="293" from-port="1" to-layer="296" to-port="0" /> + <edge from-layer="294" from-port="0" to-layer="296" to-port="1" /> + <edge from-layer="295" from-port="0" to-layer="296" to-port="2" /> + <edge from-layer="296" from-port="3" to-layer="298" to-port="0" /> + <edge from-layer="297" from-port="0" to-layer="298" to-port="1" /> + <edge from-layer="298" from-port="2" to-layer="301" to-port="1" /> + <edge from-layer="299" from-port="0" to-layer="301" to-port="2" /> + <edge from-layer="300" from-port="0" to-layer="301" to-port="3" /> + <edge from-layer="301" from-port="4" to-layer="302" to-port="1" /> + <edge from-layer="302" from-port="2" to-layer="304" to-port="0" /> + <edge from-layer="303" from-port="0" to-layer="304" to-port="1" /> + <edge from-layer="304" from-port="2" to-layer="305" to-port="1" /> + <edge from-layer="305" from-port="2" to-layer="307" to-port="0" /> + <edge from-layer="306" from-port="0" to-layer="307" to-port="1" /> + <edge from-layer="307" from-port="2" to-layer="308" to-port="0" /> + <edge from-layer="308" from-port="2" to-layer="309" to-port="0" /> + <edge from-layer="309" from-port="1" to-layer="332" to-port="0" /> + <edge from-layer="310" from-port="0" to-layer="313" to-port="0" /> + <edge from-layer="311" from-port="0" to-layer="312" to-port="1" /> + <edge from-layer="312" from-port="2" to-layer="313" to-port="1" /> + <edge from-layer="313" from-port="2" to-layer="329" to-port="0" /> + <edge from-layer="313" from-port="2" to-layer="320" to-port="0" /> + <edge from-layer="313" from-port="2" to-layer="314" to-port="0" /> + <edge from-layer="314" from-port="1" to-layer="317" to-port="0" /> + <edge from-layer="315" from-port="0" to-layer="317" to-port="1" /> + <edge from-layer="316" from-port="0" to-layer="317" to-port="2" /> + <edge from-layer="317" from-port="3" to-layer="319" to-port="0" /> + <edge from-layer="318" from-port="0" to-layer="319" to-port="1" /> + <edge from-layer="319" from-port="2" to-layer="328" to-port="0" /> + <edge from-layer="320" from-port="1" to-layer="323" to-port="0" /> + <edge from-layer="321" from-port="0" to-layer="323" to-port="1" /> + <edge from-layer="322" from-port="0" to-layer="323" to-port="2" /> + <edge from-layer="323" from-port="3" to-layer="325" to-port="0" /> + <edge from-layer="324" from-port="0" to-layer="325" to-port="1" /> + <edge from-layer="325" from-port="2" to-layer="328" to-port="1" /> + <edge from-layer="326" from-port="0" to-layer="328" to-port="2" /> + <edge from-layer="327" from-port="0" to-layer="328" to-port="3" /> + <edge from-layer="328" from-port="4" to-layer="329" to-port="1" /> + <edge from-layer="329" from-port="2" to-layer="331" to-port="0" /> + <edge from-layer="330" from-port="0" to-layer="331" to-port="1" /> + <edge from-layer="331" from-port="2" to-layer="332" to-port="1" /> + <edge from-layer="332" from-port="2" to-layer="334" to-port="0" /> + <edge from-layer="333" from-port="0" to-layer="334" to-port="1" /> + <edge from-layer="334" from-port="2" to-layer="335" to-port="0" /> + <edge from-layer="334" from-port="2" to-layer="341" to-port="0" /> + <edge from-layer="334" from-port="2" to-layer="349" to-port="0" /> + <edge from-layer="335" from-port="1" to-layer="338" to-port="0" /> + <edge from-layer="336" from-port="0" to-layer="338" to-port="1" /> + <edge from-layer="337" from-port="0" to-layer="338" to-port="2" /> + <edge from-layer="338" from-port="3" to-layer="340" to-port="0" /> + <edge from-layer="339" from-port="0" to-layer="340" to-port="1" /> + <edge from-layer="340" from-port="2" to-layer="348" to-port="0" /> + <edge from-layer="341" from-port="1" to-layer="344" to-port="0" /> + <edge from-layer="342" from-port="0" to-layer="344" to-port="1" /> + <edge from-layer="343" from-port="0" to-layer="344" to-port="2" /> + <edge from-layer="344" from-port="3" to-layer="346" to-port="0" /> + <edge from-layer="345" from-port="0" to-layer="346" to-port="1" /> + <edge from-layer="346" from-port="2" to-layer="348" to-port="1" /> + <edge from-layer="347" from-port="0" to-layer="348" to-port="2" /> + <edge from-layer="348" from-port="3" to-layer="349" to-port="1" /> + <edge from-layer="349" from-port="2" to-layer="351" to-port="0" /> + <edge from-layer="350" from-port="0" to-layer="351" to-port="1" /> + <edge from-layer="351" from-port="2" to-layer="352" to-port="1" /> + <edge from-layer="352" from-port="2" to-layer="353" to-port="0" /> + <edge from-layer="353" from-port="2" to-layer="356" to-port="0" /> + <edge from-layer="353" from-port="2" to-layer="355" to-port="0" /> + <edge from-layer="354" from-port="0" to-layer="355" to-port="1" /> + <edge from-layer="355" from-port="2" to-layer="356" to-port="1" /> + <edge from-layer="356" from-port="2" to-layer="364" to-port="0" /> + <edge from-layer="356" from-port="2" to-layer="358" to-port="0" /> + <edge from-layer="357" from-port="0" to-layer="358" to-port="1" /> + <edge from-layer="358" from-port="2" to-layer="360" to-port="0" /> + <edge from-layer="359" from-port="0" to-layer="360" to-port="1" /> + <edge from-layer="360" from-port="2" to-layer="362" to-port="0" /> + <edge from-layer="361" from-port="0" to-layer="362" to-port="1" /> + <edge from-layer="362" from-port="2" to-layer="363" to-port="0" /> + <edge from-layer="363" from-port="1" to-layer="364" to-port="1" /> + <edge from-layer="364" from-port="2" to-layer="366" to-port="0" /> + <edge from-layer="365" from-port="0" to-layer="366" to-port="1" /> + <edge from-layer="366" from-port="2" to-layer="368" to-port="0" /> + <edge from-layer="367" from-port="0" to-layer="368" to-port="1" /> + <edge from-layer="368" from-port="2" to-layer="370" to-port="0" /> + <edge from-layer="368" from-port="2" to-layer="383" to-port="1" /> + <edge from-layer="369" from-port="0" to-layer="370" to-port="1" /> + <edge from-layer="370" from-port="2" to-layer="371" to-port="1" /> + <edge from-layer="371" from-port="2" to-layer="377" to-port="0" /> + <edge from-layer="371" from-port="2" to-layer="373" to-port="0" /> + <edge from-layer="372" from-port="0" to-layer="373" to-port="1" /> + <edge from-layer="373" from-port="2" to-layer="374" to-port="0" /> + <edge from-layer="374" from-port="1" to-layer="376" to-port="0" /> + <edge from-layer="375" from-port="0" to-layer="376" to-port="1" /> + <edge from-layer="376" from-port="2" to-layer="377" to-port="1" /> + <edge from-layer="377" from-port="2" to-layer="379" to-port="0" /> + <edge from-layer="378" from-port="0" to-layer="379" to-port="1" /> + <edge from-layer="379" from-port="2" to-layer="381" to-port="0" /> + <edge from-layer="380" from-port="0" to-layer="381" to-port="1" /> + <edge from-layer="381" from-port="2" to-layer="382" to-port="1" /> + <edge from-layer="382" from-port="2" to-layer="383" to-port="0" /> + <edge from-layer="383" from-port="2" to-layer="386" to-port="0" /> + <edge from-layer="383" from-port="2" to-layer="385" to-port="0" /> + <edge from-layer="384" from-port="0" to-layer="385" to-port="1" /> + <edge from-layer="385" from-port="2" to-layer="386" to-port="1" /> + <edge from-layer="386" from-port="2" to-layer="388" to-port="0" /> + <edge from-layer="386" from-port="2" to-layer="394" to-port="0" /> + <edge from-layer="387" from-port="0" to-layer="388" to-port="1" /> + <edge from-layer="388" from-port="2" to-layer="390" to-port="0" /> + <edge from-layer="389" from-port="0" to-layer="390" to-port="1" /> + <edge from-layer="390" from-port="2" to-layer="392" to-port="0" /> + <edge from-layer="391" from-port="0" to-layer="392" to-port="1" /> + <edge from-layer="392" from-port="2" to-layer="393" to-port="0" /> + <edge from-layer="393" from-port="1" to-layer="394" to-port="1" /> + <edge from-layer="394" from-port="2" to-layer="396" to-port="0" /> + <edge from-layer="395" from-port="0" to-layer="396" to-port="1" /> + <edge from-layer="396" from-port="2" to-layer="398" to-port="0" /> + <edge from-layer="397" from-port="0" to-layer="398" to-port="1" /> + <edge from-layer="398" from-port="2" to-layer="449" to-port="0" /> + <edge from-layer="398" from-port="2" to-layer="490" to-port="1" /> + <edge from-layer="398" from-port="2" to-layer="422" to-port="0" /> + <edge from-layer="398" from-port="2" to-layer="400" to-port="0" /> + <edge from-layer="399" from-port="0" to-layer="400" to-port="1" /> + <edge from-layer="400" from-port="2" to-layer="401" to-port="1" /> + <edge from-layer="401" from-port="2" to-layer="417" to-port="0" /> + <edge from-layer="401" from-port="2" to-layer="402" to-port="0" /> + <edge from-layer="401" from-port="2" to-layer="408" to-port="0" /> + <edge from-layer="402" from-port="1" to-layer="405" to-port="0" /> + <edge from-layer="403" from-port="0" to-layer="405" to-port="1" /> + <edge from-layer="404" from-port="0" to-layer="405" to-port="2" /> + <edge from-layer="405" from-port="3" to-layer="407" to-port="0" /> + <edge from-layer="406" from-port="0" to-layer="407" to-port="1" /> + <edge from-layer="407" from-port="2" to-layer="416" to-port="0" /> + <edge from-layer="408" from-port="1" to-layer="411" to-port="0" /> + <edge from-layer="409" from-port="0" to-layer="411" to-port="1" /> + <edge from-layer="410" from-port="0" to-layer="411" to-port="2" /> + <edge from-layer="411" from-port="3" to-layer="413" to-port="0" /> + <edge from-layer="412" from-port="0" to-layer="413" to-port="1" /> + <edge from-layer="413" from-port="2" to-layer="416" to-port="1" /> + <edge from-layer="414" from-port="0" to-layer="416" to-port="2" /> + <edge from-layer="415" from-port="0" to-layer="416" to-port="3" /> + <edge from-layer="416" from-port="4" to-layer="417" to-port="1" /> + <edge from-layer="417" from-port="2" to-layer="419" to-port="0" /> + <edge from-layer="418" from-port="0" to-layer="419" to-port="1" /> + <edge from-layer="419" from-port="2" to-layer="442" to-port="0" /> + <edge from-layer="420" from-port="0" to-layer="423" to-port="0" /> + <edge from-layer="421" from-port="0" to-layer="422" to-port="1" /> + <edge from-layer="422" from-port="2" to-layer="423" to-port="1" /> + <edge from-layer="423" from-port="2" to-layer="430" to-port="0" /> + <edge from-layer="423" from-port="2" to-layer="439" to-port="0" /> + <edge from-layer="423" from-port="2" to-layer="424" to-port="0" /> + <edge from-layer="424" from-port="1" to-layer="427" to-port="0" /> + <edge from-layer="425" from-port="0" to-layer="427" to-port="1" /> + <edge from-layer="426" from-port="0" to-layer="427" to-port="2" /> + <edge from-layer="427" from-port="3" to-layer="429" to-port="0" /> + <edge from-layer="428" from-port="0" to-layer="429" to-port="1" /> + <edge from-layer="429" from-port="2" to-layer="438" to-port="0" /> + <edge from-layer="430" from-port="1" to-layer="433" to-port="0" /> + <edge from-layer="431" from-port="0" to-layer="433" to-port="1" /> + <edge from-layer="432" from-port="0" to-layer="433" to-port="2" /> + <edge from-layer="433" from-port="3" to-layer="435" to-port="0" /> + <edge from-layer="434" from-port="0" to-layer="435" to-port="1" /> + <edge from-layer="435" from-port="2" to-layer="438" to-port="1" /> + <edge from-layer="436" from-port="0" to-layer="438" to-port="2" /> + <edge from-layer="437" from-port="0" to-layer="438" to-port="3" /> + <edge from-layer="438" from-port="4" to-layer="439" to-port="1" /> + <edge from-layer="439" from-port="2" to-layer="441" to-port="0" /> + <edge from-layer="440" from-port="0" to-layer="441" to-port="1" /> + <edge from-layer="441" from-port="2" to-layer="442" to-port="1" /> + <edge from-layer="442" from-port="2" to-layer="444" to-port="0" /> + <edge from-layer="443" from-port="0" to-layer="444" to-port="1" /> + <edge from-layer="444" from-port="2" to-layer="445" to-port="0" /> + <edge from-layer="445" from-port="2" to-layer="446" to-port="0" /> + <edge from-layer="446" from-port="1" to-layer="469" to-port="0" /> + <edge from-layer="447" from-port="0" to-layer="450" to-port="0" /> + <edge from-layer="448" from-port="0" to-layer="449" to-port="1" /> + <edge from-layer="449" from-port="2" to-layer="450" to-port="1" /> + <edge from-layer="450" from-port="2" to-layer="457" to-port="0" /> + <edge from-layer="450" from-port="2" to-layer="451" to-port="0" /> + <edge from-layer="450" from-port="2" to-layer="466" to-port="0" /> + <edge from-layer="451" from-port="1" to-layer="454" to-port="0" /> + <edge from-layer="452" from-port="0" to-layer="454" to-port="1" /> + <edge from-layer="453" from-port="0" to-layer="454" to-port="2" /> + <edge from-layer="454" from-port="3" to-layer="456" to-port="0" /> + <edge from-layer="455" from-port="0" to-layer="456" to-port="1" /> + <edge from-layer="456" from-port="2" to-layer="465" to-port="0" /> + <edge from-layer="457" from-port="1" to-layer="460" to-port="0" /> + <edge from-layer="458" from-port="0" to-layer="460" to-port="1" /> + <edge from-layer="459" from-port="0" to-layer="460" to-port="2" /> + <edge from-layer="460" from-port="3" to-layer="462" to-port="0" /> + <edge from-layer="461" from-port="0" to-layer="462" to-port="1" /> + <edge from-layer="462" from-port="2" to-layer="465" to-port="1" /> + <edge from-layer="463" from-port="0" to-layer="465" to-port="2" /> + <edge from-layer="464" from-port="0" to-layer="465" to-port="3" /> + <edge from-layer="465" from-port="4" to-layer="466" to-port="1" /> + <edge from-layer="466" from-port="2" to-layer="468" to-port="0" /> + <edge from-layer="467" from-port="0" to-layer="468" to-port="1" /> + <edge from-layer="468" from-port="2" to-layer="469" to-port="1" /> + <edge from-layer="469" from-port="2" to-layer="471" to-port="0" /> + <edge from-layer="470" from-port="0" to-layer="471" to-port="1" /> + <edge from-layer="471" from-port="2" to-layer="486" to-port="0" /> + <edge from-layer="471" from-port="2" to-layer="472" to-port="0" /> + <edge from-layer="471" from-port="2" to-layer="478" to-port="0" /> + <edge from-layer="472" from-port="1" to-layer="475" to-port="0" /> + <edge from-layer="473" from-port="0" to-layer="475" to-port="1" /> + <edge from-layer="474" from-port="0" to-layer="475" to-port="2" /> + <edge from-layer="475" from-port="3" to-layer="477" to-port="0" /> + <edge from-layer="476" from-port="0" to-layer="477" to-port="1" /> + <edge from-layer="477" from-port="2" to-layer="485" to-port="0" /> + <edge from-layer="478" from-port="1" to-layer="481" to-port="0" /> + <edge from-layer="479" from-port="0" to-layer="481" to-port="1" /> + <edge from-layer="480" from-port="0" to-layer="481" to-port="2" /> + <edge from-layer="481" from-port="3" to-layer="483" to-port="0" /> + <edge from-layer="482" from-port="0" to-layer="483" to-port="1" /> + <edge from-layer="483" from-port="2" to-layer="485" to-port="1" /> + <edge from-layer="484" from-port="0" to-layer="485" to-port="2" /> + <edge from-layer="485" from-port="3" to-layer="486" to-port="1" /> + <edge from-layer="486" from-port="2" to-layer="488" to-port="0" /> + <edge from-layer="487" from-port="0" to-layer="488" to-port="1" /> + <edge from-layer="488" from-port="2" to-layer="489" to-port="1" /> + <edge from-layer="489" from-port="2" to-layer="490" to-port="0" /> + <edge from-layer="490" from-port="2" to-layer="493" to-port="0" /> + <edge from-layer="490" from-port="2" to-layer="492" to-port="0" /> + <edge from-layer="491" from-port="0" to-layer="492" to-port="1" /> + <edge from-layer="492" from-port="2" to-layer="493" to-port="1" /> + <edge from-layer="493" from-port="2" to-layer="495" to-port="0" /> + <edge from-layer="493" from-port="2" to-layer="501" to-port="0" /> + <edge from-layer="494" from-port="0" to-layer="495" to-port="1" /> + <edge from-layer="495" from-port="2" to-layer="497" to-port="0" /> + <edge from-layer="496" from-port="0" to-layer="497" to-port="1" /> + <edge from-layer="497" from-port="2" to-layer="499" to-port="0" /> + <edge from-layer="498" from-port="0" to-layer="499" to-port="1" /> + <edge from-layer="499" from-port="2" to-layer="500" to-port="0" /> + <edge from-layer="500" from-port="1" to-layer="501" to-port="1" /> + <edge from-layer="501" from-port="2" to-layer="503" to-port="0" /> + <edge from-layer="502" from-port="0" to-layer="503" to-port="1" /> + <edge from-layer="503" from-port="2" to-layer="505" to-port="0" /> + <edge from-layer="504" from-port="0" to-layer="505" to-port="1" /> + <edge from-layer="505" from-port="2" to-layer="507" to-port="0" /> + <edge from-layer="505" from-port="2" to-layer="520" to-port="1" /> + <edge from-layer="506" from-port="0" to-layer="507" to-port="1" /> + <edge from-layer="507" from-port="2" to-layer="508" to-port="1" /> + <edge from-layer="508" from-port="2" to-layer="514" to-port="0" /> + <edge from-layer="508" from-port="2" to-layer="510" to-port="0" /> + <edge from-layer="509" from-port="0" to-layer="510" to-port="1" /> + <edge from-layer="510" from-port="2" to-layer="511" to-port="0" /> + <edge from-layer="511" from-port="1" to-layer="513" to-port="0" /> + <edge from-layer="512" from-port="0" to-layer="513" to-port="1" /> + <edge from-layer="513" from-port="2" to-layer="514" to-port="1" /> + <edge from-layer="514" from-port="2" to-layer="516" to-port="0" /> + <edge from-layer="515" from-port="0" to-layer="516" to-port="1" /> + <edge from-layer="516" from-port="2" to-layer="518" to-port="0" /> + <edge from-layer="517" from-port="0" to-layer="518" to-port="1" /> + <edge from-layer="518" from-port="2" to-layer="519" to-port="1" /> + <edge from-layer="519" from-port="2" to-layer="520" to-port="0" /> + <edge from-layer="520" from-port="2" to-layer="522" to-port="0" /> + <edge from-layer="520" from-port="2" to-layer="523" to-port="0" /> + <edge from-layer="521" from-port="0" to-layer="522" to-port="1" /> + <edge from-layer="522" from-port="2" to-layer="523" to-port="1" /> + <edge from-layer="523" from-port="2" to-layer="531" to-port="0" /> + <edge from-layer="523" from-port="2" to-layer="525" to-port="0" /> + <edge from-layer="524" from-port="0" to-layer="525" to-port="1" /> + <edge from-layer="525" from-port="2" to-layer="527" to-port="0" /> + <edge from-layer="526" from-port="0" to-layer="527" to-port="1" /> + <edge from-layer="527" from-port="2" to-layer="529" to-port="0" /> + <edge from-layer="528" from-port="0" to-layer="529" to-port="1" /> + <edge from-layer="529" from-port="2" to-layer="530" to-port="0" /> + <edge from-layer="530" from-port="1" to-layer="531" to-port="1" /> + <edge from-layer="531" from-port="2" to-layer="533" to-port="0" /> + <edge from-layer="532" from-port="0" to-layer="533" to-port="1" /> + <edge from-layer="533" from-port="2" to-layer="535" to-port="0" /> + <edge from-layer="534" from-port="0" to-layer="535" to-port="1" /> + <edge from-layer="535" from-port="2" to-layer="537" to-port="0" /> + <edge from-layer="535" from-port="2" to-layer="627" to-port="1" /> + <edge from-layer="535" from-port="2" to-layer="559" to-port="0" /> + <edge from-layer="535" from-port="2" to-layer="586" to-port="0" /> + <edge from-layer="536" from-port="0" to-layer="537" to-port="1" /> + <edge from-layer="537" from-port="2" to-layer="538" to-port="1" /> + <edge from-layer="538" from-port="2" to-layer="539" to-port="0" /> + <edge from-layer="538" from-port="2" to-layer="554" to-port="0" /> + <edge from-layer="538" from-port="2" to-layer="545" to-port="0" /> + <edge from-layer="539" from-port="1" to-layer="542" to-port="0" /> + <edge from-layer="540" from-port="0" to-layer="542" to-port="1" /> + <edge from-layer="541" from-port="0" to-layer="542" to-port="2" /> + <edge from-layer="542" from-port="3" to-layer="544" to-port="0" /> + <edge from-layer="543" from-port="0" to-layer="544" to-port="1" /> + <edge from-layer="544" from-port="2" to-layer="553" to-port="0" /> + <edge from-layer="545" from-port="1" to-layer="548" to-port="0" /> + <edge from-layer="546" from-port="0" to-layer="548" to-port="1" /> + <edge from-layer="547" from-port="0" to-layer="548" to-port="2" /> + <edge from-layer="548" from-port="3" to-layer="550" to-port="0" /> + <edge from-layer="549" from-port="0" to-layer="550" to-port="1" /> + <edge from-layer="550" from-port="2" to-layer="553" to-port="1" /> + <edge from-layer="551" from-port="0" to-layer="553" to-port="2" /> + <edge from-layer="552" from-port="0" to-layer="553" to-port="3" /> + <edge from-layer="553" from-port="4" to-layer="554" to-port="1" /> + <edge from-layer="554" from-port="2" to-layer="556" to-port="0" /> + <edge from-layer="555" from-port="0" to-layer="556" to-port="1" /> + <edge from-layer="556" from-port="2" to-layer="579" to-port="0" /> + <edge from-layer="557" from-port="0" to-layer="560" to-port="0" /> + <edge from-layer="558" from-port="0" to-layer="559" to-port="1" /> + <edge from-layer="559" from-port="2" to-layer="560" to-port="1" /> + <edge from-layer="560" from-port="2" to-layer="561" to-port="0" /> + <edge from-layer="560" from-port="2" to-layer="567" to-port="0" /> + <edge from-layer="560" from-port="2" to-layer="576" to-port="0" /> + <edge from-layer="561" from-port="1" to-layer="564" to-port="0" /> + <edge from-layer="562" from-port="0" to-layer="564" to-port="1" /> + <edge from-layer="563" from-port="0" to-layer="564" to-port="2" /> + <edge from-layer="564" from-port="3" to-layer="566" to-port="0" /> + <edge from-layer="565" from-port="0" to-layer="566" to-port="1" /> + <edge from-layer="566" from-port="2" to-layer="575" to-port="0" /> + <edge from-layer="567" from-port="1" to-layer="570" to-port="0" /> + <edge from-layer="568" from-port="0" to-layer="570" to-port="1" /> + <edge from-layer="569" from-port="0" to-layer="570" to-port="2" /> + <edge from-layer="570" from-port="3" to-layer="572" to-port="0" /> + <edge from-layer="571" from-port="0" to-layer="572" to-port="1" /> + <edge from-layer="572" from-port="2" to-layer="575" to-port="1" /> + <edge from-layer="573" from-port="0" to-layer="575" to-port="2" /> + <edge from-layer="574" from-port="0" to-layer="575" to-port="3" /> + <edge from-layer="575" from-port="4" to-layer="576" to-port="1" /> + <edge from-layer="576" from-port="2" to-layer="578" to-port="0" /> + <edge from-layer="577" from-port="0" to-layer="578" to-port="1" /> + <edge from-layer="578" from-port="2" to-layer="579" to-port="1" /> + <edge from-layer="579" from-port="2" to-layer="581" to-port="0" /> + <edge from-layer="580" from-port="0" to-layer="581" to-port="1" /> + <edge from-layer="581" from-port="2" to-layer="582" to-port="0" /> + <edge from-layer="582" from-port="2" to-layer="583" to-port="0" /> + <edge from-layer="583" from-port="1" to-layer="606" to-port="0" /> + <edge from-layer="584" from-port="0" to-layer="587" to-port="0" /> + <edge from-layer="585" from-port="0" to-layer="586" to-port="1" /> + <edge from-layer="586" from-port="2" to-layer="587" to-port="1" /> + <edge from-layer="587" from-port="2" to-layer="594" to-port="0" /> + <edge from-layer="587" from-port="2" to-layer="588" to-port="0" /> + <edge from-layer="587" from-port="2" to-layer="603" to-port="0" /> + <edge from-layer="588" from-port="1" to-layer="591" to-port="0" /> + <edge from-layer="589" from-port="0" to-layer="591" to-port="1" /> + <edge from-layer="590" from-port="0" to-layer="591" to-port="2" /> + <edge from-layer="591" from-port="3" to-layer="593" to-port="0" /> + <edge from-layer="592" from-port="0" to-layer="593" to-port="1" /> + <edge from-layer="593" from-port="2" to-layer="602" to-port="0" /> + <edge from-layer="594" from-port="1" to-layer="597" to-port="0" /> + <edge from-layer="595" from-port="0" to-layer="597" to-port="1" /> + <edge from-layer="596" from-port="0" to-layer="597" to-port="2" /> + <edge from-layer="597" from-port="3" to-layer="599" to-port="0" /> + <edge from-layer="598" from-port="0" to-layer="599" to-port="1" /> + <edge from-layer="599" from-port="2" to-layer="602" to-port="1" /> + <edge from-layer="600" from-port="0" to-layer="602" to-port="2" /> + <edge from-layer="601" from-port="0" to-layer="602" to-port="3" /> + <edge from-layer="602" from-port="4" to-layer="603" to-port="1" /> + <edge from-layer="603" from-port="2" to-layer="605" to-port="0" /> + <edge from-layer="604" from-port="0" to-layer="605" to-port="1" /> + <edge from-layer="605" from-port="2" to-layer="606" to-port="1" /> + <edge from-layer="606" from-port="2" to-layer="608" to-port="0" /> + <edge from-layer="607" from-port="0" to-layer="608" to-port="1" /> + <edge from-layer="608" from-port="2" to-layer="609" to-port="0" /> + <edge from-layer="608" from-port="2" to-layer="623" to-port="0" /> + <edge from-layer="608" from-port="2" to-layer="615" to-port="0" /> + <edge from-layer="609" from-port="1" to-layer="612" to-port="0" /> + <edge from-layer="610" from-port="0" to-layer="612" to-port="1" /> + <edge from-layer="611" from-port="0" to-layer="612" to-port="2" /> + <edge from-layer="612" from-port="3" to-layer="614" to-port="0" /> + <edge from-layer="613" from-port="0" to-layer="614" to-port="1" /> + <edge from-layer="614" from-port="2" to-layer="622" to-port="0" /> + <edge from-layer="615" from-port="1" to-layer="618" to-port="0" /> + <edge from-layer="616" from-port="0" to-layer="618" to-port="1" /> + <edge from-layer="617" from-port="0" to-layer="618" to-port="2" /> + <edge from-layer="618" from-port="3" to-layer="620" to-port="0" /> + <edge from-layer="619" from-port="0" to-layer="620" to-port="1" /> + <edge from-layer="620" from-port="2" to-layer="622" to-port="1" /> + <edge from-layer="621" from-port="0" to-layer="622" to-port="2" /> + <edge from-layer="622" from-port="3" to-layer="623" to-port="1" /> + <edge from-layer="623" from-port="2" to-layer="625" to-port="0" /> + <edge from-layer="624" from-port="0" to-layer="625" to-port="1" /> + <edge from-layer="625" from-port="2" to-layer="626" to-port="1" /> + <edge from-layer="626" from-port="2" to-layer="627" to-port="0" /> + <edge from-layer="627" from-port="2" to-layer="630" to-port="0" /> + <edge from-layer="627" from-port="2" to-layer="629" to-port="0" /> + <edge from-layer="628" from-port="0" to-layer="629" to-port="1" /> + <edge from-layer="629" from-port="2" to-layer="630" to-port="1" /> + <edge from-layer="630" from-port="2" to-layer="632" to-port="0" /> + <edge from-layer="630" from-port="2" to-layer="638" to-port="0" /> + <edge from-layer="631" from-port="0" to-layer="632" to-port="1" /> + <edge from-layer="632" from-port="2" to-layer="634" to-port="0" /> + <edge from-layer="633" from-port="0" to-layer="634" to-port="1" /> + <edge from-layer="634" from-port="2" to-layer="636" to-port="0" /> + <edge from-layer="635" from-port="0" to-layer="636" to-port="1" /> + <edge from-layer="636" from-port="2" to-layer="637" to-port="0" /> + <edge from-layer="637" from-port="1" to-layer="638" to-port="1" /> + <edge from-layer="638" from-port="2" to-layer="640" to-port="0" /> + <edge from-layer="639" from-port="0" to-layer="640" to-port="1" /> + <edge from-layer="640" from-port="2" to-layer="642" to-port="0" /> + <edge from-layer="641" from-port="0" to-layer="642" to-port="1" /> + <edge from-layer="642" from-port="2" to-layer="657" to-port="1" /> + <edge from-layer="642" from-port="2" to-layer="644" to-port="0" /> + <edge from-layer="643" from-port="0" to-layer="644" to-port="1" /> + <edge from-layer="644" from-port="2" to-layer="645" to-port="1" /> + <edge from-layer="645" from-port="2" to-layer="647" to-port="0" /> + <edge from-layer="645" from-port="2" to-layer="651" to-port="0" /> + <edge from-layer="646" from-port="0" to-layer="647" to-port="1" /> + <edge from-layer="647" from-port="2" to-layer="648" to-port="0" /> + <edge from-layer="648" from-port="1" to-layer="650" to-port="0" /> + <edge from-layer="649" from-port="0" to-layer="650" to-port="1" /> + <edge from-layer="650" from-port="2" to-layer="651" to-port="1" /> + <edge from-layer="651" from-port="2" to-layer="653" to-port="0" /> + <edge from-layer="652" from-port="0" to-layer="653" to-port="1" /> + <edge from-layer="653" from-port="2" to-layer="655" to-port="0" /> + <edge from-layer="654" from-port="0" to-layer="655" to-port="1" /> + <edge from-layer="655" from-port="2" to-layer="656" to-port="1" /> + <edge from-layer="656" from-port="2" to-layer="657" to-port="0" /> + <edge from-layer="657" from-port="2" to-layer="660" to-port="0" /> + <edge from-layer="657" from-port="2" to-layer="659" to-port="0" /> + <edge from-layer="658" from-port="0" to-layer="659" to-port="1" /> + <edge from-layer="659" from-port="2" to-layer="660" to-port="1" /> + <edge from-layer="660" from-port="2" to-layer="668" to-port="0" /> + <edge from-layer="660" from-port="2" to-layer="662" to-port="0" /> + <edge from-layer="661" from-port="0" to-layer="662" to-port="1" /> + <edge from-layer="662" from-port="2" to-layer="664" to-port="0" /> + <edge from-layer="663" from-port="0" to-layer="664" to-port="1" /> + <edge from-layer="664" from-port="2" to-layer="666" to-port="0" /> + <edge from-layer="665" from-port="0" to-layer="666" to-port="1" /> + <edge from-layer="666" from-port="2" to-layer="667" to-port="0" /> + <edge from-layer="667" from-port="1" to-layer="668" to-port="1" /> + <edge from-layer="668" from-port="2" to-layer="670" to-port="0" /> + <edge from-layer="669" from-port="0" to-layer="670" to-port="1" /> + <edge from-layer="670" from-port="2" to-layer="672" to-port="0" /> + <edge from-layer="671" from-port="0" to-layer="672" to-port="1" /> + <edge from-layer="672" from-port="2" to-layer="764" to-port="1" /> + <edge from-layer="672" from-port="2" to-layer="674" to-port="0" /> + <edge from-layer="672" from-port="2" to-layer="696" to-port="0" /> + <edge from-layer="672" from-port="2" to-layer="723" to-port="0" /> + <edge from-layer="673" from-port="0" to-layer="674" to-port="1" /> + <edge from-layer="674" from-port="2" to-layer="675" to-port="1" /> + <edge from-layer="675" from-port="2" to-layer="691" to-port="0" /> + <edge from-layer="675" from-port="2" to-layer="682" to-port="0" /> + <edge from-layer="675" from-port="2" to-layer="676" to-port="0" /> + <edge from-layer="676" from-port="1" to-layer="679" to-port="0" /> + <edge from-layer="677" from-port="0" to-layer="679" to-port="1" /> + <edge from-layer="678" from-port="0" to-layer="679" to-port="2" /> + <edge from-layer="679" from-port="3" to-layer="681" to-port="0" /> + <edge from-layer="680" from-port="0" to-layer="681" to-port="1" /> + <edge from-layer="681" from-port="2" to-layer="690" to-port="0" /> + <edge from-layer="682" from-port="1" to-layer="685" to-port="0" /> + <edge from-layer="683" from-port="0" to-layer="685" to-port="1" /> + <edge from-layer="684" from-port="0" to-layer="685" to-port="2" /> + <edge from-layer="685" from-port="3" to-layer="687" to-port="0" /> + <edge from-layer="686" from-port="0" to-layer="687" to-port="1" /> + <edge from-layer="687" from-port="2" to-layer="690" to-port="1" /> + <edge from-layer="688" from-port="0" to-layer="690" to-port="2" /> + <edge from-layer="689" from-port="0" to-layer="690" to-port="3" /> + <edge from-layer="690" from-port="4" to-layer="691" to-port="1" /> + <edge from-layer="691" from-port="2" to-layer="693" to-port="0" /> + <edge from-layer="692" from-port="0" to-layer="693" to-port="1" /> + <edge from-layer="693" from-port="2" to-layer="716" to-port="0" /> + <edge from-layer="694" from-port="0" to-layer="697" to-port="0" /> + <edge from-layer="695" from-port="0" to-layer="696" to-port="1" /> + <edge from-layer="696" from-port="2" to-layer="697" to-port="1" /> + <edge from-layer="697" from-port="2" to-layer="698" to-port="0" /> + <edge from-layer="697" from-port="2" to-layer="704" to-port="0" /> + <edge from-layer="697" from-port="2" to-layer="713" to-port="0" /> + <edge from-layer="698" from-port="1" to-layer="701" to-port="0" /> + <edge from-layer="699" from-port="0" to-layer="701" to-port="1" /> + <edge from-layer="700" from-port="0" to-layer="701" to-port="2" /> + <edge from-layer="701" from-port="3" to-layer="703" to-port="0" /> + <edge from-layer="702" from-port="0" to-layer="703" to-port="1" /> + <edge from-layer="703" from-port="2" to-layer="712" to-port="0" /> + <edge from-layer="704" from-port="1" to-layer="707" to-port="0" /> + <edge from-layer="705" from-port="0" to-layer="707" to-port="1" /> + <edge from-layer="706" from-port="0" to-layer="707" to-port="2" /> + <edge from-layer="707" from-port="3" to-layer="709" to-port="0" /> + <edge from-layer="708" from-port="0" to-layer="709" to-port="1" /> + <edge from-layer="709" from-port="2" to-layer="712" to-port="1" /> + <edge from-layer="710" from-port="0" to-layer="712" to-port="2" /> + <edge from-layer="711" from-port="0" to-layer="712" to-port="3" /> + <edge from-layer="712" from-port="4" to-layer="713" to-port="1" /> + <edge from-layer="713" from-port="2" to-layer="715" to-port="0" /> + <edge from-layer="714" from-port="0" to-layer="715" to-port="1" /> + <edge from-layer="715" from-port="2" to-layer="716" to-port="1" /> + <edge from-layer="716" from-port="2" to-layer="718" to-port="0" /> + <edge from-layer="717" from-port="0" to-layer="718" to-port="1" /> + <edge from-layer="718" from-port="2" to-layer="719" to-port="0" /> + <edge from-layer="719" from-port="2" to-layer="720" to-port="0" /> + <edge from-layer="720" from-port="1" to-layer="743" to-port="0" /> + <edge from-layer="721" from-port="0" to-layer="724" to-port="0" /> + <edge from-layer="722" from-port="0" to-layer="723" to-port="1" /> + <edge from-layer="723" from-port="2" to-layer="724" to-port="1" /> + <edge from-layer="724" from-port="2" to-layer="731" to-port="0" /> + <edge from-layer="724" from-port="2" to-layer="725" to-port="0" /> + <edge from-layer="724" from-port="2" to-layer="740" to-port="0" /> + <edge from-layer="725" from-port="1" to-layer="728" to-port="0" /> + <edge from-layer="726" from-port="0" to-layer="728" to-port="1" /> + <edge from-layer="727" from-port="0" to-layer="728" to-port="2" /> + <edge from-layer="728" from-port="3" to-layer="730" to-port="0" /> + <edge from-layer="729" from-port="0" to-layer="730" to-port="1" /> + <edge from-layer="730" from-port="2" to-layer="739" to-port="0" /> + <edge from-layer="731" from-port="1" to-layer="734" to-port="0" /> + <edge from-layer="732" from-port="0" to-layer="734" to-port="1" /> + <edge from-layer="733" from-port="0" to-layer="734" to-port="2" /> + <edge from-layer="734" from-port="3" to-layer="736" to-port="0" /> + <edge from-layer="735" from-port="0" to-layer="736" to-port="1" /> + <edge from-layer="736" from-port="2" to-layer="739" to-port="1" /> + <edge from-layer="737" from-port="0" to-layer="739" to-port="2" /> + <edge from-layer="738" from-port="0" to-layer="739" to-port="3" /> + <edge from-layer="739" from-port="4" to-layer="740" to-port="1" /> + <edge from-layer="740" from-port="2" to-layer="742" to-port="0" /> + <edge from-layer="741" from-port="0" to-layer="742" to-port="1" /> + <edge from-layer="742" from-port="2" to-layer="743" to-port="1" /> + <edge from-layer="743" from-port="2" to-layer="745" to-port="0" /> + <edge from-layer="744" from-port="0" to-layer="745" to-port="1" /> + <edge from-layer="745" from-port="2" to-layer="760" to-port="0" /> + <edge from-layer="745" from-port="2" to-layer="752" to-port="0" /> + <edge from-layer="745" from-port="2" to-layer="746" to-port="0" /> + <edge from-layer="746" from-port="1" to-layer="749" to-port="0" /> + <edge from-layer="747" from-port="0" to-layer="749" to-port="1" /> + <edge from-layer="748" from-port="0" to-layer="749" to-port="2" /> + <edge from-layer="749" from-port="3" to-layer="751" to-port="0" /> + <edge from-layer="750" from-port="0" to-layer="751" to-port="1" /> + <edge from-layer="751" from-port="2" to-layer="759" to-port="0" /> + <edge from-layer="752" from-port="1" to-layer="755" to-port="0" /> + <edge from-layer="753" from-port="0" to-layer="755" to-port="1" /> + <edge from-layer="754" from-port="0" to-layer="755" to-port="2" /> + <edge from-layer="755" from-port="3" to-layer="757" to-port="0" /> + <edge from-layer="756" from-port="0" to-layer="757" to-port="1" /> + <edge from-layer="757" from-port="2" to-layer="759" to-port="1" /> + <edge from-layer="758" from-port="0" to-layer="759" to-port="2" /> + <edge from-layer="759" from-port="3" to-layer="760" to-port="1" /> + <edge from-layer="760" from-port="2" to-layer="762" to-port="0" /> + <edge from-layer="761" from-port="0" to-layer="762" to-port="1" /> + <edge from-layer="762" from-port="2" to-layer="763" to-port="1" /> + <edge from-layer="763" from-port="2" to-layer="764" to-port="0" /> + <edge from-layer="764" from-port="2" to-layer="767" to-port="0" /> + <edge from-layer="764" from-port="2" to-layer="766" to-port="0" /> + <edge from-layer="765" from-port="0" to-layer="766" to-port="1" /> + <edge from-layer="766" from-port="2" to-layer="767" to-port="1" /> + <edge from-layer="767" from-port="2" to-layer="769" to-port="0" /> + <edge from-layer="767" from-port="2" to-layer="775" to-port="0" /> + <edge from-layer="768" from-port="0" to-layer="769" to-port="1" /> + <edge from-layer="769" from-port="2" to-layer="771" to-port="0" /> + <edge from-layer="770" from-port="0" to-layer="771" to-port="1" /> + <edge from-layer="771" from-port="2" to-layer="773" to-port="0" /> + <edge from-layer="772" from-port="0" to-layer="773" to-port="1" /> + <edge from-layer="773" from-port="2" to-layer="774" to-port="0" /> + <edge from-layer="774" from-port="1" to-layer="775" to-port="1" /> + <edge from-layer="775" from-port="2" to-layer="777" to-port="0" /> + <edge from-layer="776" from-port="0" to-layer="777" to-port="1" /> + <edge from-layer="777" from-port="2" to-layer="779" to-port="0" /> + <edge from-layer="778" from-port="0" to-layer="779" to-port="1" /> + <edge from-layer="779" from-port="2" to-layer="794" to-port="1" /> + <edge from-layer="779" from-port="2" to-layer="781" to-port="0" /> + <edge from-layer="780" from-port="0" to-layer="781" to-port="1" /> + <edge from-layer="781" from-port="2" to-layer="782" to-port="1" /> + <edge from-layer="782" from-port="2" to-layer="784" to-port="0" /> + <edge from-layer="782" from-port="2" to-layer="788" to-port="0" /> + <edge from-layer="783" from-port="0" to-layer="784" to-port="1" /> + <edge from-layer="784" from-port="2" to-layer="785" to-port="0" /> + <edge from-layer="785" from-port="1" to-layer="787" to-port="0" /> + <edge from-layer="786" from-port="0" to-layer="787" to-port="1" /> + <edge from-layer="787" from-port="2" to-layer="788" to-port="1" /> + <edge from-layer="788" from-port="2" to-layer="790" to-port="0" /> + <edge from-layer="789" from-port="0" to-layer="790" to-port="1" /> + <edge from-layer="790" from-port="2" to-layer="792" to-port="0" /> + <edge from-layer="791" from-port="0" to-layer="792" to-port="1" /> + <edge from-layer="792" from-port="2" to-layer="793" to-port="1" /> + <edge from-layer="793" from-port="2" to-layer="794" to-port="0" /> + <edge from-layer="794" from-port="2" to-layer="797" to-port="0" /> + <edge from-layer="794" from-port="2" to-layer="796" to-port="0" /> + <edge from-layer="795" from-port="0" to-layer="796" to-port="1" /> + <edge from-layer="796" from-port="2" to-layer="797" to-port="1" /> + <edge from-layer="797" from-port="2" to-layer="805" to-port="0" /> + <edge from-layer="797" from-port="2" to-layer="799" to-port="0" /> + <edge from-layer="798" from-port="0" to-layer="799" to-port="1" /> + <edge from-layer="799" from-port="2" to-layer="801" to-port="0" /> + <edge from-layer="800" from-port="0" to-layer="801" to-port="1" /> + <edge from-layer="801" from-port="2" to-layer="803" to-port="0" /> + <edge from-layer="802" from-port="0" to-layer="803" to-port="1" /> + <edge from-layer="803" from-port="2" to-layer="804" to-port="0" /> + <edge from-layer="804" from-port="1" to-layer="805" to-port="1" /> + <edge from-layer="805" from-port="2" to-layer="807" to-port="0" /> + <edge from-layer="806" from-port="0" to-layer="807" to-port="1" /> + <edge from-layer="807" from-port="2" to-layer="809" to-port="0" /> + <edge from-layer="808" from-port="0" to-layer="809" to-port="1" /> + <edge from-layer="809" from-port="2" to-layer="860" to-port="0" /> + <edge from-layer="809" from-port="2" to-layer="901" to-port="1" /> + <edge from-layer="809" from-port="2" to-layer="811" to-port="0" /> + <edge from-layer="809" from-port="2" to-layer="833" to-port="0" /> + <edge from-layer="810" from-port="0" to-layer="811" to-port="1" /> + <edge from-layer="811" from-port="2" to-layer="812" to-port="1" /> + <edge from-layer="812" from-port="2" to-layer="828" to-port="0" /> + <edge from-layer="812" from-port="2" to-layer="813" to-port="0" /> + <edge from-layer="812" from-port="2" to-layer="819" to-port="0" /> + <edge from-layer="813" from-port="1" to-layer="816" to-port="0" /> + <edge from-layer="814" from-port="0" to-layer="816" to-port="1" /> + <edge from-layer="815" from-port="0" to-layer="816" to-port="2" /> + <edge from-layer="816" from-port="3" to-layer="818" to-port="0" /> + <edge from-layer="817" from-port="0" to-layer="818" to-port="1" /> + <edge from-layer="818" from-port="2" to-layer="827" to-port="0" /> + <edge from-layer="819" from-port="1" to-layer="822" to-port="0" /> + <edge from-layer="820" from-port="0" to-layer="822" to-port="1" /> + <edge from-layer="821" from-port="0" to-layer="822" to-port="2" /> + <edge from-layer="822" from-port="3" to-layer="824" to-port="0" /> + <edge from-layer="823" from-port="0" to-layer="824" to-port="1" /> + <edge from-layer="824" from-port="2" to-layer="827" to-port="1" /> + <edge from-layer="825" from-port="0" to-layer="827" to-port="2" /> + <edge from-layer="826" from-port="0" to-layer="827" to-port="3" /> + <edge from-layer="827" from-port="4" to-layer="828" to-port="1" /> + <edge from-layer="828" from-port="2" to-layer="830" to-port="0" /> + <edge from-layer="829" from-port="0" to-layer="830" to-port="1" /> + <edge from-layer="830" from-port="2" to-layer="853" to-port="0" /> + <edge from-layer="831" from-port="0" to-layer="834" to-port="0" /> + <edge from-layer="832" from-port="0" to-layer="833" to-port="1" /> + <edge from-layer="833" from-port="2" to-layer="834" to-port="1" /> + <edge from-layer="834" from-port="2" to-layer="841" to-port="0" /> + <edge from-layer="834" from-port="2" to-layer="850" to-port="0" /> + <edge from-layer="834" from-port="2" to-layer="835" to-port="0" /> + <edge from-layer="835" from-port="1" to-layer="838" to-port="0" /> + <edge from-layer="836" from-port="0" to-layer="838" to-port="1" /> + <edge from-layer="837" from-port="0" to-layer="838" to-port="2" /> + <edge from-layer="838" from-port="3" to-layer="840" to-port="0" /> + <edge from-layer="839" from-port="0" to-layer="840" to-port="1" /> + <edge from-layer="840" from-port="2" to-layer="849" to-port="0" /> + <edge from-layer="841" from-port="1" to-layer="844" to-port="0" /> + <edge from-layer="842" from-port="0" to-layer="844" to-port="1" /> + <edge from-layer="843" from-port="0" to-layer="844" to-port="2" /> + <edge from-layer="844" from-port="3" to-layer="846" to-port="0" /> + <edge from-layer="845" from-port="0" to-layer="846" to-port="1" /> + <edge from-layer="846" from-port="2" to-layer="849" to-port="1" /> + <edge from-layer="847" from-port="0" to-layer="849" to-port="2" /> + <edge from-layer="848" from-port="0" to-layer="849" to-port="3" /> + <edge from-layer="849" from-port="4" to-layer="850" to-port="1" /> + <edge from-layer="850" from-port="2" to-layer="852" to-port="0" /> + <edge from-layer="851" from-port="0" to-layer="852" to-port="1" /> + <edge from-layer="852" from-port="2" to-layer="853" to-port="1" /> + <edge from-layer="853" from-port="2" to-layer="855" to-port="0" /> + <edge from-layer="854" from-port="0" to-layer="855" to-port="1" /> + <edge from-layer="855" from-port="2" to-layer="856" to-port="0" /> + <edge from-layer="856" from-port="2" to-layer="857" to-port="0" /> + <edge from-layer="857" from-port="1" to-layer="880" to-port="0" /> + <edge from-layer="858" from-port="0" to-layer="861" to-port="0" /> + <edge from-layer="859" from-port="0" to-layer="860" to-port="1" /> + <edge from-layer="860" from-port="2" to-layer="861" to-port="1" /> + <edge from-layer="861" from-port="2" to-layer="862" to-port="0" /> + <edge from-layer="861" from-port="2" to-layer="868" to-port="0" /> + <edge from-layer="861" from-port="2" to-layer="877" to-port="0" /> + <edge from-layer="862" from-port="1" to-layer="865" to-port="0" /> + <edge from-layer="863" from-port="0" to-layer="865" to-port="1" /> + <edge from-layer="864" from-port="0" to-layer="865" to-port="2" /> + <edge from-layer="865" from-port="3" to-layer="867" to-port="0" /> + <edge from-layer="866" from-port="0" to-layer="867" to-port="1" /> + <edge from-layer="867" from-port="2" to-layer="876" to-port="0" /> + <edge from-layer="868" from-port="1" to-layer="871" to-port="0" /> + <edge from-layer="869" from-port="0" to-layer="871" to-port="1" /> + <edge from-layer="870" from-port="0" to-layer="871" to-port="2" /> + <edge from-layer="871" from-port="3" to-layer="873" to-port="0" /> + <edge from-layer="872" from-port="0" to-layer="873" to-port="1" /> + <edge from-layer="873" from-port="2" to-layer="876" to-port="1" /> + <edge from-layer="874" from-port="0" to-layer="876" to-port="2" /> + <edge from-layer="875" from-port="0" to-layer="876" to-port="3" /> + <edge from-layer="876" from-port="4" to-layer="877" to-port="1" /> + <edge from-layer="877" from-port="2" to-layer="879" to-port="0" /> + <edge from-layer="878" from-port="0" to-layer="879" to-port="1" /> + <edge from-layer="879" from-port="2" to-layer="880" to-port="1" /> + <edge from-layer="880" from-port="2" to-layer="882" to-port="0" /> + <edge from-layer="881" from-port="0" to-layer="882" to-port="1" /> + <edge from-layer="882" from-port="2" to-layer="883" to-port="0" /> + <edge from-layer="882" from-port="2" to-layer="897" to-port="0" /> + <edge from-layer="882" from-port="2" to-layer="889" to-port="0" /> + <edge from-layer="883" from-port="1" to-layer="886" to-port="0" /> + <edge from-layer="884" from-port="0" to-layer="886" to-port="1" /> + <edge from-layer="885" from-port="0" to-layer="886" to-port="2" /> + <edge from-layer="886" from-port="3" to-layer="888" to-port="0" /> + <edge from-layer="887" from-port="0" to-layer="888" to-port="1" /> + <edge from-layer="888" from-port="2" to-layer="896" to-port="0" /> + <edge from-layer="889" from-port="1" to-layer="892" to-port="0" /> + <edge from-layer="890" from-port="0" to-layer="892" to-port="1" /> + <edge from-layer="891" from-port="0" to-layer="892" to-port="2" /> + <edge from-layer="892" from-port="3" to-layer="894" to-port="0" /> + <edge from-layer="893" from-port="0" to-layer="894" to-port="1" /> + <edge from-layer="894" from-port="2" to-layer="896" to-port="1" /> + <edge from-layer="895" from-port="0" to-layer="896" to-port="2" /> + <edge from-layer="896" from-port="3" to-layer="897" to-port="1" /> + <edge from-layer="897" from-port="2" to-layer="899" to-port="0" /> + <edge from-layer="898" from-port="0" to-layer="899" to-port="1" /> + <edge from-layer="899" from-port="2" to-layer="900" to-port="1" /> + <edge from-layer="900" from-port="2" to-layer="901" to-port="0" /> + <edge from-layer="901" from-port="2" to-layer="903" to-port="0" /> + <edge from-layer="901" from-port="2" to-layer="904" to-port="0" /> + <edge from-layer="902" from-port="0" to-layer="903" to-port="1" /> + <edge from-layer="903" from-port="2" to-layer="904" to-port="1" /> + <edge from-layer="904" from-port="2" to-layer="912" to-port="0" /> + <edge from-layer="904" from-port="2" to-layer="906" to-port="0" /> + <edge from-layer="905" from-port="0" to-layer="906" to-port="1" /> + <edge from-layer="906" from-port="2" to-layer="908" to-port="0" /> + <edge from-layer="907" from-port="0" to-layer="908" to-port="1" /> + <edge from-layer="908" from-port="2" to-layer="910" to-port="0" /> + <edge from-layer="909" from-port="0" to-layer="910" to-port="1" /> + <edge from-layer="910" from-port="2" to-layer="911" to-port="0" /> + <edge from-layer="911" from-port="1" to-layer="912" to-port="1" /> + <edge from-layer="912" from-port="2" to-layer="914" to-port="0" /> + <edge from-layer="913" from-port="0" to-layer="914" to-port="1" /> + <edge from-layer="914" from-port="2" to-layer="916" to-port="0" /> + <edge from-layer="915" from-port="0" to-layer="916" to-port="1" /> + <edge from-layer="916" from-port="2" to-layer="931" to-port="1" /> + <edge from-layer="916" from-port="2" to-layer="918" to-port="0" /> + <edge from-layer="917" from-port="0" to-layer="918" to-port="1" /> + <edge from-layer="918" from-port="2" to-layer="919" to-port="1" /> + <edge from-layer="919" from-port="2" to-layer="921" to-port="0" /> + <edge from-layer="919" from-port="2" to-layer="925" to-port="0" /> + <edge from-layer="920" from-port="0" to-layer="921" to-port="1" /> + <edge from-layer="921" from-port="2" to-layer="922" to-port="0" /> + <edge from-layer="922" from-port="1" to-layer="924" to-port="0" /> + <edge from-layer="923" from-port="0" to-layer="924" to-port="1" /> + <edge from-layer="924" from-port="2" to-layer="925" to-port="1" /> + <edge from-layer="925" from-port="2" to-layer="927" to-port="0" /> + <edge from-layer="926" from-port="0" to-layer="927" to-port="1" /> + <edge from-layer="927" from-port="2" to-layer="929" to-port="0" /> + <edge from-layer="928" from-port="0" to-layer="929" to-port="1" /> + <edge from-layer="929" from-port="2" to-layer="930" to-port="1" /> + <edge from-layer="930" from-port="2" to-layer="931" to-port="0" /> + <edge from-layer="931" from-port="2" to-layer="933" to-port="0" /> + <edge from-layer="931" from-port="2" to-layer="934" to-port="0" /> + <edge from-layer="932" from-port="0" to-layer="933" to-port="1" /> + <edge from-layer="933" from-port="2" to-layer="934" to-port="1" /> + <edge from-layer="934" from-port="2" to-layer="936" to-port="0" /> + <edge from-layer="934" from-port="2" to-layer="942" to-port="0" /> + <edge from-layer="935" from-port="0" to-layer="936" to-port="1" /> + <edge from-layer="936" from-port="2" to-layer="938" to-port="0" /> + <edge from-layer="937" from-port="0" to-layer="938" to-port="1" /> + <edge from-layer="938" from-port="2" to-layer="940" to-port="0" /> + <edge from-layer="939" from-port="0" to-layer="940" to-port="1" /> + <edge from-layer="940" from-port="2" to-layer="941" to-port="0" /> + <edge from-layer="941" from-port="1" to-layer="942" to-port="1" /> + <edge from-layer="942" from-port="2" to-layer="944" to-port="0" /> + <edge from-layer="943" from-port="0" to-layer="944" to-port="1" /> + <edge from-layer="944" from-port="2" to-layer="946" to-port="0" /> + <edge from-layer="945" from-port="0" to-layer="946" to-port="1" /> + <edge from-layer="946" from-port="2" to-layer="948" to-port="0" /> + <edge from-layer="946" from-port="2" to-layer="1038" to-port="1" /> + <edge from-layer="946" from-port="2" to-layer="997" to-port="0" /> + <edge from-layer="946" from-port="2" to-layer="970" to-port="0" /> + <edge from-layer="947" from-port="0" to-layer="948" to-port="1" /> + <edge from-layer="948" from-port="2" to-layer="949" to-port="1" /> + <edge from-layer="949" from-port="2" to-layer="956" to-port="0" /> + <edge from-layer="949" from-port="2" to-layer="965" to-port="0" /> + <edge from-layer="949" from-port="2" to-layer="950" to-port="0" /> + <edge from-layer="950" from-port="1" to-layer="953" to-port="0" /> + <edge from-layer="951" from-port="0" to-layer="953" to-port="1" /> + <edge from-layer="952" from-port="0" to-layer="953" to-port="2" /> + <edge from-layer="953" from-port="3" to-layer="955" to-port="0" /> + <edge from-layer="954" from-port="0" to-layer="955" to-port="1" /> + <edge from-layer="955" from-port="2" to-layer="964" to-port="0" /> + <edge from-layer="956" from-port="1" to-layer="959" to-port="0" /> + <edge from-layer="957" from-port="0" to-layer="959" to-port="1" /> + <edge from-layer="958" from-port="0" to-layer="959" to-port="2" /> + <edge from-layer="959" from-port="3" to-layer="961" to-port="0" /> + <edge from-layer="960" from-port="0" to-layer="961" to-port="1" /> + <edge from-layer="961" from-port="2" to-layer="964" to-port="1" /> + <edge from-layer="962" from-port="0" to-layer="964" to-port="2" /> + <edge from-layer="963" from-port="0" to-layer="964" to-port="3" /> + <edge from-layer="964" from-port="4" to-layer="965" to-port="1" /> + <edge from-layer="965" from-port="2" to-layer="967" to-port="0" /> + <edge from-layer="966" from-port="0" to-layer="967" to-port="1" /> + <edge from-layer="967" from-port="2" to-layer="990" to-port="0" /> + <edge from-layer="968" from-port="0" to-layer="971" to-port="0" /> + <edge from-layer="969" from-port="0" to-layer="970" to-port="1" /> + <edge from-layer="970" from-port="2" to-layer="971" to-port="1" /> + <edge from-layer="971" from-port="2" to-layer="978" to-port="0" /> + <edge from-layer="971" from-port="2" to-layer="972" to-port="0" /> + <edge from-layer="971" from-port="2" to-layer="987" to-port="0" /> + <edge from-layer="972" from-port="1" to-layer="975" to-port="0" /> + <edge from-layer="973" from-port="0" to-layer="975" to-port="1" /> + <edge from-layer="974" from-port="0" to-layer="975" to-port="2" /> + <edge from-layer="975" from-port="3" to-layer="977" to-port="0" /> + <edge from-layer="976" from-port="0" to-layer="977" to-port="1" /> + <edge from-layer="977" from-port="2" to-layer="986" to-port="0" /> + <edge from-layer="978" from-port="1" to-layer="981" to-port="0" /> + <edge from-layer="979" from-port="0" to-layer="981" to-port="1" /> + <edge from-layer="980" from-port="0" to-layer="981" to-port="2" /> + <edge from-layer="981" from-port="3" to-layer="983" to-port="0" /> + <edge from-layer="982" from-port="0" to-layer="983" to-port="1" /> + <edge from-layer="983" from-port="2" to-layer="986" to-port="1" /> + <edge from-layer="984" from-port="0" to-layer="986" to-port="2" /> + <edge from-layer="985" from-port="0" to-layer="986" to-port="3" /> + <edge from-layer="986" from-port="4" to-layer="987" to-port="1" /> + <edge from-layer="987" from-port="2" to-layer="989" to-port="0" /> + <edge from-layer="988" from-port="0" to-layer="989" to-port="1" /> + <edge from-layer="989" from-port="2" to-layer="990" to-port="1" /> + <edge from-layer="990" from-port="2" to-layer="992" to-port="0" /> + <edge from-layer="991" from-port="0" to-layer="992" to-port="1" /> + <edge from-layer="992" from-port="2" to-layer="993" to-port="0" /> + <edge from-layer="993" from-port="2" to-layer="994" to-port="0" /> + <edge from-layer="994" from-port="1" to-layer="1017" to-port="0" /> + <edge from-layer="995" from-port="0" to-layer="998" to-port="0" /> + <edge from-layer="996" from-port="0" to-layer="997" to-port="1" /> + <edge from-layer="997" from-port="2" to-layer="998" to-port="1" /> + <edge from-layer="998" from-port="2" to-layer="1014" to-port="0" /> + <edge from-layer="998" from-port="2" to-layer="1005" to-port="0" /> + <edge from-layer="998" from-port="2" to-layer="999" to-port="0" /> + <edge from-layer="999" from-port="1" to-layer="1002" to-port="0" /> + <edge from-layer="1000" from-port="0" to-layer="1002" to-port="1" /> + <edge from-layer="1001" from-port="0" to-layer="1002" to-port="2" /> + <edge from-layer="1002" from-port="3" to-layer="1004" to-port="0" /> + <edge from-layer="1003" from-port="0" to-layer="1004" to-port="1" /> + <edge from-layer="1004" from-port="2" to-layer="1013" to-port="0" /> + <edge from-layer="1005" from-port="1" to-layer="1008" to-port="0" /> + <edge from-layer="1006" from-port="0" to-layer="1008" to-port="1" /> + <edge from-layer="1007" from-port="0" to-layer="1008" to-port="2" /> + <edge from-layer="1008" from-port="3" to-layer="1010" to-port="0" /> + <edge from-layer="1009" from-port="0" to-layer="1010" to-port="1" /> + <edge from-layer="1010" from-port="2" to-layer="1013" to-port="1" /> + <edge from-layer="1011" from-port="0" to-layer="1013" to-port="2" /> + <edge from-layer="1012" from-port="0" to-layer="1013" to-port="3" /> + <edge from-layer="1013" from-port="4" to-layer="1014" to-port="1" /> + <edge from-layer="1014" from-port="2" to-layer="1016" to-port="0" /> + <edge from-layer="1015" from-port="0" to-layer="1016" to-port="1" /> + <edge from-layer="1016" from-port="2" to-layer="1017" to-port="1" /> + <edge from-layer="1017" from-port="2" to-layer="1019" to-port="0" /> + <edge from-layer="1018" from-port="0" to-layer="1019" to-port="1" /> + <edge from-layer="1019" from-port="2" to-layer="1020" to-port="0" /> + <edge from-layer="1019" from-port="2" to-layer="1026" to-port="0" /> + <edge from-layer="1019" from-port="2" to-layer="1034" to-port="0" /> + <edge from-layer="1020" from-port="1" to-layer="1023" to-port="0" /> + <edge from-layer="1021" from-port="0" to-layer="1023" to-port="1" /> + <edge from-layer="1022" from-port="0" to-layer="1023" to-port="2" /> + <edge from-layer="1023" from-port="3" to-layer="1025" to-port="0" /> + <edge from-layer="1024" from-port="0" to-layer="1025" to-port="1" /> + <edge from-layer="1025" from-port="2" to-layer="1033" to-port="0" /> + <edge from-layer="1026" from-port="1" to-layer="1029" to-port="0" /> + <edge from-layer="1027" from-port="0" to-layer="1029" to-port="1" /> + <edge from-layer="1028" from-port="0" to-layer="1029" to-port="2" /> + <edge from-layer="1029" from-port="3" to-layer="1031" to-port="0" /> + <edge from-layer="1030" from-port="0" to-layer="1031" to-port="1" /> + <edge from-layer="1031" from-port="2" to-layer="1033" to-port="1" /> + <edge from-layer="1032" from-port="0" to-layer="1033" to-port="2" /> + <edge from-layer="1033" from-port="3" to-layer="1034" to-port="1" /> + <edge from-layer="1034" from-port="2" to-layer="1036" to-port="0" /> + <edge from-layer="1035" from-port="0" to-layer="1036" to-port="1" /> + <edge from-layer="1036" from-port="2" to-layer="1037" to-port="1" /> + <edge from-layer="1037" from-port="2" to-layer="1038" to-port="0" /> + <edge from-layer="1038" from-port="2" to-layer="1040" to-port="0" /> + <edge from-layer="1038" from-port="2" to-layer="1041" to-port="0" /> + <edge from-layer="1039" from-port="0" to-layer="1040" to-port="1" /> + <edge from-layer="1040" from-port="2" to-layer="1041" to-port="1" /> + <edge from-layer="1041" from-port="2" to-layer="1049" to-port="0" /> + <edge from-layer="1041" from-port="2" to-layer="1043" to-port="0" /> + <edge from-layer="1042" from-port="0" to-layer="1043" to-port="1" /> + <edge from-layer="1043" from-port="2" to-layer="1045" to-port="0" /> + <edge from-layer="1044" from-port="0" to-layer="1045" to-port="1" /> + <edge from-layer="1045" from-port="2" to-layer="1047" to-port="0" /> + <edge from-layer="1046" from-port="0" to-layer="1047" to-port="1" /> + <edge from-layer="1047" from-port="2" to-layer="1048" to-port="0" /> + <edge from-layer="1048" from-port="1" to-layer="1049" to-port="1" /> + <edge from-layer="1049" from-port="2" to-layer="1051" to-port="0" /> + <edge from-layer="1050" from-port="0" to-layer="1051" to-port="1" /> + <edge from-layer="1051" from-port="2" to-layer="1053" to-port="0" /> + <edge from-layer="1052" from-port="0" to-layer="1053" to-port="1" /> + <edge from-layer="1053" from-port="2" to-layer="1068" to-port="1" /> + <edge from-layer="1053" from-port="2" to-layer="1055" to-port="0" /> + <edge from-layer="1054" from-port="0" to-layer="1055" to-port="1" /> + <edge from-layer="1055" from-port="2" to-layer="1056" to-port="1" /> + <edge from-layer="1056" from-port="2" to-layer="1062" to-port="0" /> + <edge from-layer="1056" from-port="2" to-layer="1058" to-port="0" /> + <edge from-layer="1057" from-port="0" to-layer="1058" to-port="1" /> + <edge from-layer="1058" from-port="2" to-layer="1059" to-port="0" /> + <edge from-layer="1059" from-port="1" to-layer="1061" to-port="0" /> + <edge from-layer="1060" from-port="0" to-layer="1061" to-port="1" /> + <edge from-layer="1061" from-port="2" to-layer="1062" to-port="1" /> + <edge from-layer="1062" from-port="2" to-layer="1064" to-port="0" /> + <edge from-layer="1063" from-port="0" to-layer="1064" to-port="1" /> + <edge from-layer="1064" from-port="2" to-layer="1066" to-port="0" /> + <edge from-layer="1065" from-port="0" to-layer="1066" to-port="1" /> + <edge from-layer="1066" from-port="2" to-layer="1067" to-port="1" /> + <edge from-layer="1067" from-port="2" to-layer="1068" to-port="0" /> + <edge from-layer="1068" from-port="2" to-layer="1071" to-port="0" /> + <edge from-layer="1068" from-port="2" to-layer="1070" to-port="0" /> + <edge from-layer="1069" from-port="0" to-layer="1070" to-port="1" /> + <edge from-layer="1070" from-port="2" to-layer="1071" to-port="1" /> + <edge from-layer="1071" from-port="2" to-layer="1079" to-port="0" /> + <edge from-layer="1071" from-port="2" to-layer="1073" to-port="0" /> + <edge from-layer="1072" from-port="0" to-layer="1073" to-port="1" /> + <edge from-layer="1073" from-port="2" to-layer="1075" to-port="0" /> + <edge from-layer="1074" from-port="0" to-layer="1075" to-port="1" /> + <edge from-layer="1075" from-port="2" to-layer="1077" to-port="0" /> + <edge from-layer="1076" from-port="0" to-layer="1077" to-port="1" /> + <edge from-layer="1077" from-port="2" to-layer="1078" to-port="0" /> + <edge from-layer="1078" from-port="1" to-layer="1079" to-port="1" /> + <edge from-layer="1079" from-port="2" to-layer="1081" to-port="0" /> + <edge from-layer="1080" from-port="0" to-layer="1081" to-port="1" /> + <edge from-layer="1081" from-port="2" to-layer="1083" to-port="0" /> + <edge from-layer="1082" from-port="0" to-layer="1083" to-port="1" /> + <edge from-layer="1083" from-port="2" to-layer="1134" to-port="0" /> + <edge from-layer="1083" from-port="2" to-layer="1175" to-port="1" /> + <edge from-layer="1083" from-port="2" to-layer="1107" to-port="0" /> + <edge from-layer="1083" from-port="2" to-layer="1085" to-port="0" /> + <edge from-layer="1084" from-port="0" to-layer="1085" to-port="1" /> + <edge from-layer="1085" from-port="2" to-layer="1086" to-port="1" /> + <edge from-layer="1086" from-port="2" to-layer="1087" to-port="0" /> + <edge from-layer="1086" from-port="2" to-layer="1093" to-port="0" /> + <edge from-layer="1086" from-port="2" to-layer="1102" to-port="0" /> + <edge from-layer="1087" from-port="1" to-layer="1090" to-port="0" /> + <edge from-layer="1088" from-port="0" to-layer="1090" to-port="1" /> + <edge from-layer="1089" from-port="0" to-layer="1090" to-port="2" /> + <edge from-layer="1090" from-port="3" to-layer="1092" to-port="0" /> + <edge from-layer="1091" from-port="0" to-layer="1092" to-port="1" /> + <edge from-layer="1092" from-port="2" to-layer="1101" to-port="0" /> + <edge from-layer="1093" from-port="1" to-layer="1096" to-port="0" /> + <edge from-layer="1094" from-port="0" to-layer="1096" to-port="1" /> + <edge from-layer="1095" from-port="0" to-layer="1096" to-port="2" /> + <edge from-layer="1096" from-port="3" to-layer="1098" to-port="0" /> + <edge from-layer="1097" from-port="0" to-layer="1098" to-port="1" /> + <edge from-layer="1098" from-port="2" to-layer="1101" to-port="1" /> + <edge from-layer="1099" from-port="0" to-layer="1101" to-port="2" /> + <edge from-layer="1100" from-port="0" to-layer="1101" to-port="3" /> + <edge from-layer="1101" from-port="4" to-layer="1102" to-port="1" /> + <edge from-layer="1102" from-port="2" to-layer="1104" to-port="0" /> + <edge from-layer="1103" from-port="0" to-layer="1104" to-port="1" /> + <edge from-layer="1104" from-port="2" to-layer="1127" to-port="0" /> + <edge from-layer="1105" from-port="0" to-layer="1108" to-port="0" /> + <edge from-layer="1106" from-port="0" to-layer="1107" to-port="1" /> + <edge from-layer="1107" from-port="2" to-layer="1108" to-port="1" /> + <edge from-layer="1108" from-port="2" to-layer="1109" to-port="0" /> + <edge from-layer="1108" from-port="2" to-layer="1115" to-port="0" /> + <edge from-layer="1108" from-port="2" to-layer="1124" to-port="0" /> + <edge from-layer="1109" from-port="1" to-layer="1112" to-port="0" /> + <edge from-layer="1110" from-port="0" to-layer="1112" to-port="1" /> + <edge from-layer="1111" from-port="0" to-layer="1112" to-port="2" /> + <edge from-layer="1112" from-port="3" to-layer="1114" to-port="0" /> + <edge from-layer="1113" from-port="0" to-layer="1114" to-port="1" /> + <edge from-layer="1114" from-port="2" to-layer="1123" to-port="0" /> + <edge from-layer="1115" from-port="1" to-layer="1118" to-port="0" /> + <edge from-layer="1116" from-port="0" to-layer="1118" to-port="1" /> + <edge from-layer="1117" from-port="0" to-layer="1118" to-port="2" /> + <edge from-layer="1118" from-port="3" to-layer="1120" to-port="0" /> + <edge from-layer="1119" from-port="0" to-layer="1120" to-port="1" /> + <edge from-layer="1120" from-port="2" to-layer="1123" to-port="1" /> + <edge from-layer="1121" from-port="0" to-layer="1123" to-port="2" /> + <edge from-layer="1122" from-port="0" to-layer="1123" to-port="3" /> + <edge from-layer="1123" from-port="4" to-layer="1124" to-port="1" /> + <edge from-layer="1124" from-port="2" to-layer="1126" to-port="0" /> + <edge from-layer="1125" from-port="0" to-layer="1126" to-port="1" /> + <edge from-layer="1126" from-port="2" to-layer="1127" to-port="1" /> + <edge from-layer="1127" from-port="2" to-layer="1129" to-port="0" /> + <edge from-layer="1128" from-port="0" to-layer="1129" to-port="1" /> + <edge from-layer="1129" from-port="2" to-layer="1130" to-port="0" /> + <edge from-layer="1130" from-port="2" to-layer="1131" to-port="0" /> + <edge from-layer="1131" from-port="1" to-layer="1154" to-port="0" /> + <edge from-layer="1132" from-port="0" to-layer="1135" to-port="0" /> + <edge from-layer="1133" from-port="0" to-layer="1134" to-port="1" /> + <edge from-layer="1134" from-port="2" to-layer="1135" to-port="1" /> + <edge from-layer="1135" from-port="2" to-layer="1142" to-port="0" /> + <edge from-layer="1135" from-port="2" to-layer="1151" to-port="0" /> + <edge from-layer="1135" from-port="2" to-layer="1136" to-port="0" /> + <edge from-layer="1136" from-port="1" to-layer="1139" to-port="0" /> + <edge from-layer="1137" from-port="0" to-layer="1139" to-port="1" /> + <edge from-layer="1138" from-port="0" to-layer="1139" to-port="2" /> + <edge from-layer="1139" from-port="3" to-layer="1141" to-port="0" /> + <edge from-layer="1140" from-port="0" to-layer="1141" to-port="1" /> + <edge from-layer="1141" from-port="2" to-layer="1150" to-port="0" /> + <edge from-layer="1142" from-port="1" to-layer="1145" to-port="0" /> + <edge from-layer="1143" from-port="0" to-layer="1145" to-port="1" /> + <edge from-layer="1144" from-port="0" to-layer="1145" to-port="2" /> + <edge from-layer="1145" from-port="3" to-layer="1147" to-port="0" /> + <edge from-layer="1146" from-port="0" to-layer="1147" to-port="1" /> + <edge from-layer="1147" from-port="2" to-layer="1150" to-port="1" /> + <edge from-layer="1148" from-port="0" to-layer="1150" to-port="2" /> + <edge from-layer="1149" from-port="0" to-layer="1150" to-port="3" /> + <edge from-layer="1150" from-port="4" to-layer="1151" to-port="1" /> + <edge from-layer="1151" from-port="2" to-layer="1153" to-port="0" /> + <edge from-layer="1152" from-port="0" to-layer="1153" to-port="1" /> + <edge from-layer="1153" from-port="2" to-layer="1154" to-port="1" /> + <edge from-layer="1154" from-port="2" to-layer="1156" to-port="0" /> + <edge from-layer="1155" from-port="0" to-layer="1156" to-port="1" /> + <edge from-layer="1156" from-port="2" to-layer="1171" to-port="0" /> + <edge from-layer="1156" from-port="2" to-layer="1163" to-port="0" /> + <edge from-layer="1156" from-port="2" to-layer="1157" to-port="0" /> + <edge from-layer="1157" from-port="1" to-layer="1160" to-port="0" /> + <edge from-layer="1158" from-port="0" to-layer="1160" to-port="1" /> + <edge from-layer="1159" from-port="0" to-layer="1160" to-port="2" /> + <edge from-layer="1160" from-port="3" to-layer="1162" to-port="0" /> + <edge from-layer="1161" from-port="0" to-layer="1162" to-port="1" /> + <edge from-layer="1162" from-port="2" to-layer="1170" to-port="0" /> + <edge from-layer="1163" from-port="1" to-layer="1166" to-port="0" /> + <edge from-layer="1164" from-port="0" to-layer="1166" to-port="1" /> + <edge from-layer="1165" from-port="0" to-layer="1166" to-port="2" /> + <edge from-layer="1166" from-port="3" to-layer="1168" to-port="0" /> + <edge from-layer="1167" from-port="0" to-layer="1168" to-port="1" /> + <edge from-layer="1168" from-port="2" to-layer="1170" to-port="1" /> + <edge from-layer="1169" from-port="0" to-layer="1170" to-port="2" /> + <edge from-layer="1170" from-port="3" to-layer="1171" to-port="1" /> + <edge from-layer="1171" from-port="2" to-layer="1173" to-port="0" /> + <edge from-layer="1172" from-port="0" to-layer="1173" to-port="1" /> + <edge from-layer="1173" from-port="2" to-layer="1174" to-port="1" /> + <edge from-layer="1174" from-port="2" to-layer="1175" to-port="0" /> + <edge from-layer="1175" from-port="2" to-layer="1178" to-port="0" /> + <edge from-layer="1175" from-port="2" to-layer="1177" to-port="0" /> + <edge from-layer="1176" from-port="0" to-layer="1177" to-port="1" /> + <edge from-layer="1177" from-port="2" to-layer="1178" to-port="1" /> + <edge from-layer="1178" from-port="2" to-layer="1180" to-port="0" /> + <edge from-layer="1178" from-port="2" to-layer="1186" to-port="0" /> + <edge from-layer="1179" from-port="0" to-layer="1180" to-port="1" /> + <edge from-layer="1180" from-port="2" to-layer="1182" to-port="0" /> + <edge from-layer="1181" from-port="0" to-layer="1182" to-port="1" /> + <edge from-layer="1182" from-port="2" to-layer="1184" to-port="0" /> + <edge from-layer="1183" from-port="0" to-layer="1184" to-port="1" /> + <edge from-layer="1184" from-port="2" to-layer="1185" to-port="0" /> + <edge from-layer="1185" from-port="1" to-layer="1186" to-port="1" /> + <edge from-layer="1186" from-port="2" to-layer="1188" to-port="0" /> + <edge from-layer="1187" from-port="0" to-layer="1188" to-port="1" /> + <edge from-layer="1188" from-port="2" to-layer="1190" to-port="0" /> + <edge from-layer="1189" from-port="0" to-layer="1190" to-port="1" /> + <edge from-layer="1190" from-port="2" to-layer="1192" to-port="0" /> + <edge from-layer="1190" from-port="2" to-layer="1205" to-port="1" /> + <edge from-layer="1191" from-port="0" to-layer="1192" to-port="1" /> + <edge from-layer="1192" from-port="2" to-layer="1193" to-port="1" /> + <edge from-layer="1193" from-port="2" to-layer="1195" to-port="0" /> + <edge from-layer="1193" from-port="2" to-layer="1199" to-port="0" /> + <edge from-layer="1194" from-port="0" to-layer="1195" to-port="1" /> + <edge from-layer="1195" from-port="2" to-layer="1196" to-port="0" /> + <edge from-layer="1196" from-port="1" to-layer="1198" to-port="0" /> + <edge from-layer="1197" from-port="0" to-layer="1198" to-port="1" /> + <edge from-layer="1198" from-port="2" to-layer="1199" to-port="1" /> + <edge from-layer="1199" from-port="2" to-layer="1201" to-port="0" /> + <edge from-layer="1200" from-port="0" to-layer="1201" to-port="1" /> + <edge from-layer="1201" from-port="2" to-layer="1203" to-port="0" /> + <edge from-layer="1202" from-port="0" to-layer="1203" to-port="1" /> + <edge from-layer="1203" from-port="2" to-layer="1204" to-port="1" /> + <edge from-layer="1204" from-port="2" to-layer="1205" to-port="0" /> + <edge from-layer="1205" from-port="2" to-layer="1208" to-port="0" /> + <edge from-layer="1205" from-port="2" to-layer="1207" to-port="0" /> + <edge from-layer="1206" from-port="0" to-layer="1207" to-port="1" /> + <edge from-layer="1207" from-port="2" to-layer="1208" to-port="1" /> + <edge from-layer="1208" from-port="2" to-layer="1216" to-port="0" /> + <edge from-layer="1208" from-port="2" to-layer="1210" to-port="0" /> + <edge from-layer="1209" from-port="0" to-layer="1210" to-port="1" /> + <edge from-layer="1210" from-port="2" to-layer="1212" to-port="0" /> + <edge from-layer="1211" from-port="0" to-layer="1212" to-port="1" /> + <edge from-layer="1212" from-port="2" to-layer="1214" to-port="0" /> + <edge from-layer="1213" from-port="0" to-layer="1214" to-port="1" /> + <edge from-layer="1214" from-port="2" to-layer="1215" to-port="0" /> + <edge from-layer="1215" from-port="1" to-layer="1216" to-port="1" /> + <edge from-layer="1216" from-port="2" to-layer="1218" to-port="0" /> + <edge from-layer="1217" from-port="0" to-layer="1218" to-port="1" /> + <edge from-layer="1218" from-port="2" to-layer="1220" to-port="0" /> + <edge from-layer="1219" from-port="0" to-layer="1220" to-port="1" /> + <edge from-layer="1220" from-port="2" to-layer="1271" to-port="0" /> + <edge from-layer="1220" from-port="2" to-layer="1312" to-port="1" /> + <edge from-layer="1220" from-port="2" to-layer="1222" to-port="0" /> + <edge from-layer="1220" from-port="2" to-layer="1244" to-port="0" /> + <edge from-layer="1221" from-port="0" to-layer="1222" to-port="1" /> + <edge from-layer="1222" from-port="2" to-layer="1223" to-port="1" /> + <edge from-layer="1223" from-port="2" to-layer="1224" to-port="0" /> + <edge from-layer="1223" from-port="2" to-layer="1239" to-port="0" /> + <edge from-layer="1223" from-port="2" to-layer="1230" to-port="0" /> + <edge from-layer="1224" from-port="1" to-layer="1227" to-port="0" /> + <edge from-layer="1225" from-port="0" to-layer="1227" to-port="1" /> + <edge from-layer="1226" from-port="0" to-layer="1227" to-port="2" /> + <edge from-layer="1227" from-port="3" to-layer="1229" to-port="0" /> + <edge from-layer="1228" from-port="0" to-layer="1229" to-port="1" /> + <edge from-layer="1229" from-port="2" to-layer="1238" to-port="0" /> + <edge from-layer="1230" from-port="1" to-layer="1233" to-port="0" /> + <edge from-layer="1231" from-port="0" to-layer="1233" to-port="1" /> + <edge from-layer="1232" from-port="0" to-layer="1233" to-port="2" /> + <edge from-layer="1233" from-port="3" to-layer="1235" to-port="0" /> + <edge from-layer="1234" from-port="0" to-layer="1235" to-port="1" /> + <edge from-layer="1235" from-port="2" to-layer="1238" to-port="1" /> + <edge from-layer="1236" from-port="0" to-layer="1238" to-port="2" /> + <edge from-layer="1237" from-port="0" to-layer="1238" to-port="3" /> + <edge from-layer="1238" from-port="4" to-layer="1239" to-port="1" /> + <edge from-layer="1239" from-port="2" to-layer="1241" to-port="0" /> + <edge from-layer="1240" from-port="0" to-layer="1241" to-port="1" /> + <edge from-layer="1241" from-port="2" to-layer="1264" to-port="0" /> + <edge from-layer="1242" from-port="0" to-layer="1245" to-port="0" /> + <edge from-layer="1243" from-port="0" to-layer="1244" to-port="1" /> + <edge from-layer="1244" from-port="2" to-layer="1245" to-port="1" /> + <edge from-layer="1245" from-port="2" to-layer="1246" to-port="0" /> + <edge from-layer="1245" from-port="2" to-layer="1261" to-port="0" /> + <edge from-layer="1245" from-port="2" to-layer="1252" to-port="0" /> + <edge from-layer="1246" from-port="1" to-layer="1249" to-port="0" /> + <edge from-layer="1247" from-port="0" to-layer="1249" to-port="1" /> + <edge from-layer="1248" from-port="0" to-layer="1249" to-port="2" /> + <edge from-layer="1249" from-port="3" to-layer="1251" to-port="0" /> + <edge from-layer="1250" from-port="0" to-layer="1251" to-port="1" /> + <edge from-layer="1251" from-port="2" to-layer="1260" to-port="0" /> + <edge from-layer="1252" from-port="1" to-layer="1255" to-port="0" /> + <edge from-layer="1253" from-port="0" to-layer="1255" to-port="1" /> + <edge from-layer="1254" from-port="0" to-layer="1255" to-port="2" /> + <edge from-layer="1255" from-port="3" to-layer="1257" to-port="0" /> + <edge from-layer="1256" from-port="0" to-layer="1257" to-port="1" /> + <edge from-layer="1257" from-port="2" to-layer="1260" to-port="1" /> + <edge from-layer="1258" from-port="0" to-layer="1260" to-port="2" /> + <edge from-layer="1259" from-port="0" to-layer="1260" to-port="3" /> + <edge from-layer="1260" from-port="4" to-layer="1261" to-port="1" /> + <edge from-layer="1261" from-port="2" to-layer="1263" to-port="0" /> + <edge from-layer="1262" from-port="0" to-layer="1263" to-port="1" /> + <edge from-layer="1263" from-port="2" to-layer="1264" to-port="1" /> + <edge from-layer="1264" from-port="2" to-layer="1266" to-port="0" /> + <edge from-layer="1265" from-port="0" to-layer="1266" to-port="1" /> + <edge from-layer="1266" from-port="2" to-layer="1267" to-port="0" /> + <edge from-layer="1267" from-port="2" to-layer="1268" to-port="0" /> + <edge from-layer="1268" from-port="1" to-layer="1291" to-port="0" /> + <edge from-layer="1269" from-port="0" to-layer="1272" to-port="0" /> + <edge from-layer="1270" from-port="0" to-layer="1271" to-port="1" /> + <edge from-layer="1271" from-port="2" to-layer="1272" to-port="1" /> + <edge from-layer="1272" from-port="2" to-layer="1288" to-port="0" /> + <edge from-layer="1272" from-port="2" to-layer="1273" to-port="0" /> + <edge from-layer="1272" from-port="2" to-layer="1279" to-port="0" /> + <edge from-layer="1273" from-port="1" to-layer="1276" to-port="0" /> + <edge from-layer="1274" from-port="0" to-layer="1276" to-port="1" /> + <edge from-layer="1275" from-port="0" to-layer="1276" to-port="2" /> + <edge from-layer="1276" from-port="3" to-layer="1278" to-port="0" /> + <edge from-layer="1277" from-port="0" to-layer="1278" to-port="1" /> + <edge from-layer="1278" from-port="2" to-layer="1287" to-port="0" /> + <edge from-layer="1279" from-port="1" to-layer="1282" to-port="0" /> + <edge from-layer="1280" from-port="0" to-layer="1282" to-port="1" /> + <edge from-layer="1281" from-port="0" to-layer="1282" to-port="2" /> + <edge from-layer="1282" from-port="3" to-layer="1284" to-port="0" /> + <edge from-layer="1283" from-port="0" to-layer="1284" to-port="1" /> + <edge from-layer="1284" from-port="2" to-layer="1287" to-port="1" /> + <edge from-layer="1285" from-port="0" to-layer="1287" to-port="2" /> + <edge from-layer="1286" from-port="0" to-layer="1287" to-port="3" /> + <edge from-layer="1287" from-port="4" to-layer="1288" to-port="1" /> + <edge from-layer="1288" from-port="2" to-layer="1290" to-port="0" /> + <edge from-layer="1289" from-port="0" to-layer="1290" to-port="1" /> + <edge from-layer="1290" from-port="2" to-layer="1291" to-port="1" /> + <edge from-layer="1291" from-port="2" to-layer="1293" to-port="0" /> + <edge from-layer="1292" from-port="0" to-layer="1293" to-port="1" /> + <edge from-layer="1293" from-port="2" to-layer="1294" to-port="0" /> + <edge from-layer="1293" from-port="2" to-layer="1308" to-port="0" /> + <edge from-layer="1293" from-port="2" to-layer="1300" to-port="0" /> + <edge from-layer="1294" from-port="1" to-layer="1297" to-port="0" /> + <edge from-layer="1295" from-port="0" to-layer="1297" to-port="1" /> + <edge from-layer="1296" from-port="0" to-layer="1297" to-port="2" /> + <edge from-layer="1297" from-port="3" to-layer="1299" to-port="0" /> + <edge from-layer="1298" from-port="0" to-layer="1299" to-port="1" /> + <edge from-layer="1299" from-port="2" to-layer="1307" to-port="0" /> + <edge from-layer="1300" from-port="1" to-layer="1303" to-port="0" /> + <edge from-layer="1301" from-port="0" to-layer="1303" to-port="1" /> + <edge from-layer="1302" from-port="0" to-layer="1303" to-port="2" /> + <edge from-layer="1303" from-port="3" to-layer="1305" to-port="0" /> + <edge from-layer="1304" from-port="0" to-layer="1305" to-port="1" /> + <edge from-layer="1305" from-port="2" to-layer="1307" to-port="1" /> + <edge from-layer="1306" from-port="0" to-layer="1307" to-port="2" /> + <edge from-layer="1307" from-port="3" to-layer="1308" to-port="1" /> + <edge from-layer="1308" from-port="2" to-layer="1310" to-port="0" /> + <edge from-layer="1309" from-port="0" to-layer="1310" to-port="1" /> + <edge from-layer="1310" from-port="2" to-layer="1311" to-port="1" /> + <edge from-layer="1311" from-port="2" to-layer="1312" to-port="0" /> + <edge from-layer="1312" from-port="2" to-layer="1314" to-port="0" /> + <edge from-layer="1312" from-port="2" to-layer="1315" to-port="0" /> + <edge from-layer="1313" from-port="0" to-layer="1314" to-port="1" /> + <edge from-layer="1314" from-port="2" to-layer="1315" to-port="1" /> + <edge from-layer="1315" from-port="2" to-layer="1323" to-port="0" /> + <edge from-layer="1315" from-port="2" to-layer="1317" to-port="0" /> + <edge from-layer="1316" from-port="0" to-layer="1317" to-port="1" /> + <edge from-layer="1317" from-port="2" to-layer="1319" to-port="0" /> + <edge from-layer="1318" from-port="0" to-layer="1319" to-port="1" /> + <edge from-layer="1319" from-port="2" to-layer="1321" to-port="0" /> + <edge from-layer="1320" from-port="0" to-layer="1321" to-port="1" /> + <edge from-layer="1321" from-port="2" to-layer="1322" to-port="0" /> + <edge from-layer="1322" from-port="1" to-layer="1323" to-port="1" /> + <edge from-layer="1323" from-port="2" to-layer="1325" to-port="0" /> + <edge from-layer="1324" from-port="0" to-layer="1325" to-port="1" /> + <edge from-layer="1325" from-port="2" to-layer="1327" to-port="0" /> + <edge from-layer="1326" from-port="0" to-layer="1327" to-port="1" /> + <edge from-layer="1327" from-port="2" to-layer="1342" to-port="1" /> + <edge from-layer="1327" from-port="2" to-layer="1329" to-port="0" /> + <edge from-layer="1328" from-port="0" to-layer="1329" to-port="1" /> + <edge from-layer="1329" from-port="2" to-layer="1330" to-port="1" /> + <edge from-layer="1330" from-port="2" to-layer="1336" to-port="0" /> + <edge from-layer="1330" from-port="2" to-layer="1332" to-port="0" /> + <edge from-layer="1331" from-port="0" to-layer="1332" to-port="1" /> + <edge from-layer="1332" from-port="2" to-layer="1333" to-port="0" /> + <edge from-layer="1333" from-port="1" to-layer="1335" to-port="0" /> + <edge from-layer="1334" from-port="0" to-layer="1335" to-port="1" /> + <edge from-layer="1335" from-port="2" to-layer="1336" to-port="1" /> + <edge from-layer="1336" from-port="2" to-layer="1338" to-port="0" /> + <edge from-layer="1337" from-port="0" to-layer="1338" to-port="1" /> + <edge from-layer="1338" from-port="2" to-layer="1340" to-port="0" /> + <edge from-layer="1339" from-port="0" to-layer="1340" to-port="1" /> + <edge from-layer="1340" from-port="2" to-layer="1341" to-port="1" /> + <edge from-layer="1341" from-port="2" to-layer="1342" to-port="0" /> + <edge from-layer="1342" from-port="2" to-layer="1344" to-port="0" /> + <edge from-layer="1342" from-port="2" to-layer="1345" to-port="0" /> + <edge from-layer="1343" from-port="0" to-layer="1344" to-port="1" /> + <edge from-layer="1344" from-port="2" to-layer="1345" to-port="1" /> + <edge from-layer="1345" from-port="2" to-layer="1347" to-port="0" /> + <edge from-layer="1345" from-port="2" to-layer="1353" to-port="0" /> + <edge from-layer="1346" from-port="0" to-layer="1347" to-port="1" /> + <edge from-layer="1347" from-port="2" to-layer="1349" to-port="0" /> + <edge from-layer="1348" from-port="0" to-layer="1349" to-port="1" /> + <edge from-layer="1349" from-port="2" to-layer="1351" to-port="0" /> + <edge from-layer="1350" from-port="0" to-layer="1351" to-port="1" /> + <edge from-layer="1351" from-port="2" to-layer="1352" to-port="0" /> + <edge from-layer="1352" from-port="1" to-layer="1353" to-port="1" /> + <edge from-layer="1353" from-port="2" to-layer="1355" to-port="0" /> + <edge from-layer="1354" from-port="0" to-layer="1355" to-port="1" /> + <edge from-layer="1355" from-port="2" to-layer="1357" to-port="0" /> + <edge from-layer="1356" from-port="0" to-layer="1357" to-port="1" /> + <edge from-layer="1357" from-port="2" to-layer="1359" to-port="0" /> + <edge from-layer="1357" from-port="2" to-layer="1381" to-port="0" /> + <edge from-layer="1357" from-port="2" to-layer="1449" to-port="1" /> + <edge from-layer="1357" from-port="2" to-layer="1408" to-port="0" /> + <edge from-layer="1358" from-port="0" to-layer="1359" to-port="1" /> + <edge from-layer="1359" from-port="2" to-layer="1360" to-port="1" /> + <edge from-layer="1360" from-port="2" to-layer="1361" to-port="0" /> + <edge from-layer="1360" from-port="2" to-layer="1376" to-port="0" /> + <edge from-layer="1360" from-port="2" to-layer="1367" to-port="0" /> + <edge from-layer="1361" from-port="1" to-layer="1364" to-port="0" /> + <edge from-layer="1362" from-port="0" to-layer="1364" to-port="1" /> + <edge from-layer="1363" from-port="0" to-layer="1364" to-port="2" /> + <edge from-layer="1364" from-port="3" to-layer="1366" to-port="0" /> + <edge from-layer="1365" from-port="0" to-layer="1366" to-port="1" /> + <edge from-layer="1366" from-port="2" to-layer="1375" to-port="0" /> + <edge from-layer="1367" from-port="1" to-layer="1370" to-port="0" /> + <edge from-layer="1368" from-port="0" to-layer="1370" to-port="1" /> + <edge from-layer="1369" from-port="0" to-layer="1370" to-port="2" /> + <edge from-layer="1370" from-port="3" to-layer="1372" to-port="0" /> + <edge from-layer="1371" from-port="0" to-layer="1372" to-port="1" /> + <edge from-layer="1372" from-port="2" to-layer="1375" to-port="1" /> + <edge from-layer="1373" from-port="0" to-layer="1375" to-port="2" /> + <edge from-layer="1374" from-port="0" to-layer="1375" to-port="3" /> + <edge from-layer="1375" from-port="4" to-layer="1376" to-port="1" /> + <edge from-layer="1376" from-port="2" to-layer="1378" to-port="0" /> + <edge from-layer="1377" from-port="0" to-layer="1378" to-port="1" /> + <edge from-layer="1378" from-port="2" to-layer="1401" to-port="0" /> + <edge from-layer="1379" from-port="0" to-layer="1382" to-port="0" /> + <edge from-layer="1380" from-port="0" to-layer="1381" to-port="1" /> + <edge from-layer="1381" from-port="2" to-layer="1382" to-port="1" /> + <edge from-layer="1382" from-port="2" to-layer="1398" to-port="0" /> + <edge from-layer="1382" from-port="2" to-layer="1389" to-port="0" /> + <edge from-layer="1382" from-port="2" to-layer="1383" to-port="0" /> + <edge from-layer="1383" from-port="1" to-layer="1386" to-port="0" /> + <edge from-layer="1384" from-port="0" to-layer="1386" to-port="1" /> + <edge from-layer="1385" from-port="0" to-layer="1386" to-port="2" /> + <edge from-layer="1386" from-port="3" to-layer="1388" to-port="0" /> + <edge from-layer="1387" from-port="0" to-layer="1388" to-port="1" /> + <edge from-layer="1388" from-port="2" to-layer="1397" to-port="0" /> + <edge from-layer="1389" from-port="1" to-layer="1392" to-port="0" /> + <edge from-layer="1390" from-port="0" to-layer="1392" to-port="1" /> + <edge from-layer="1391" from-port="0" to-layer="1392" to-port="2" /> + <edge from-layer="1392" from-port="3" to-layer="1394" to-port="0" /> + <edge from-layer="1393" from-port="0" to-layer="1394" to-port="1" /> + <edge from-layer="1394" from-port="2" to-layer="1397" to-port="1" /> + <edge from-layer="1395" from-port="0" to-layer="1397" to-port="2" /> + <edge from-layer="1396" from-port="0" to-layer="1397" to-port="3" /> + <edge from-layer="1397" from-port="4" to-layer="1398" to-port="1" /> + <edge from-layer="1398" from-port="2" to-layer="1400" to-port="0" /> + <edge from-layer="1399" from-port="0" to-layer="1400" to-port="1" /> + <edge from-layer="1400" from-port="2" to-layer="1401" to-port="1" /> + <edge from-layer="1401" from-port="2" to-layer="1403" to-port="0" /> + <edge from-layer="1402" from-port="0" to-layer="1403" to-port="1" /> + <edge from-layer="1403" from-port="2" to-layer="1404" to-port="0" /> + <edge from-layer="1404" from-port="2" to-layer="1405" to-port="0" /> + <edge from-layer="1405" from-port="1" to-layer="1428" to-port="0" /> + <edge from-layer="1406" from-port="0" to-layer="1409" to-port="0" /> + <edge from-layer="1407" from-port="0" to-layer="1408" to-port="1" /> + <edge from-layer="1408" from-port="2" to-layer="1409" to-port="1" /> + <edge from-layer="1409" from-port="2" to-layer="1425" to-port="0" /> + <edge from-layer="1409" from-port="2" to-layer="1410" to-port="0" /> + <edge from-layer="1409" from-port="2" to-layer="1416" to-port="0" /> + <edge from-layer="1410" from-port="1" to-layer="1413" to-port="0" /> + <edge from-layer="1411" from-port="0" to-layer="1413" to-port="1" /> + <edge from-layer="1412" from-port="0" to-layer="1413" to-port="2" /> + <edge from-layer="1413" from-port="3" to-layer="1415" to-port="0" /> + <edge from-layer="1414" from-port="0" to-layer="1415" to-port="1" /> + <edge from-layer="1415" from-port="2" to-layer="1424" to-port="0" /> + <edge from-layer="1416" from-port="1" to-layer="1419" to-port="0" /> + <edge from-layer="1417" from-port="0" to-layer="1419" to-port="1" /> + <edge from-layer="1418" from-port="0" to-layer="1419" to-port="2" /> + <edge from-layer="1419" from-port="3" to-layer="1421" to-port="0" /> + <edge from-layer="1420" from-port="0" to-layer="1421" to-port="1" /> + <edge from-layer="1421" from-port="2" to-layer="1424" to-port="1" /> + <edge from-layer="1422" from-port="0" to-layer="1424" to-port="2" /> + <edge from-layer="1423" from-port="0" to-layer="1424" to-port="3" /> + <edge from-layer="1424" from-port="4" to-layer="1425" to-port="1" /> + <edge from-layer="1425" from-port="2" to-layer="1427" to-port="0" /> + <edge from-layer="1426" from-port="0" to-layer="1427" to-port="1" /> + <edge from-layer="1427" from-port="2" to-layer="1428" to-port="1" /> + <edge from-layer="1428" from-port="2" to-layer="1430" to-port="0" /> + <edge from-layer="1429" from-port="0" to-layer="1430" to-port="1" /> + <edge from-layer="1430" from-port="2" to-layer="1431" to-port="0" /> + <edge from-layer="1430" from-port="2" to-layer="1437" to-port="0" /> + <edge from-layer="1430" from-port="2" to-layer="1445" to-port="0" /> + <edge from-layer="1431" from-port="1" to-layer="1434" to-port="0" /> + <edge from-layer="1432" from-port="0" to-layer="1434" to-port="1" /> + <edge from-layer="1433" from-port="0" to-layer="1434" to-port="2" /> + <edge from-layer="1434" from-port="3" to-layer="1436" to-port="0" /> + <edge from-layer="1435" from-port="0" to-layer="1436" to-port="1" /> + <edge from-layer="1436" from-port="2" to-layer="1444" to-port="0" /> + <edge from-layer="1437" from-port="1" to-layer="1440" to-port="0" /> + <edge from-layer="1438" from-port="0" to-layer="1440" to-port="1" /> + <edge from-layer="1439" from-port="0" to-layer="1440" to-port="2" /> + <edge from-layer="1440" from-port="3" to-layer="1442" to-port="0" /> + <edge from-layer="1441" from-port="0" to-layer="1442" to-port="1" /> + <edge from-layer="1442" from-port="2" to-layer="1444" to-port="1" /> + <edge from-layer="1443" from-port="0" to-layer="1444" to-port="2" /> + <edge from-layer="1444" from-port="3" to-layer="1445" to-port="1" /> + <edge from-layer="1445" from-port="2" to-layer="1447" to-port="0" /> + <edge from-layer="1446" from-port="0" to-layer="1447" to-port="1" /> + <edge from-layer="1447" from-port="2" to-layer="1448" to-port="1" /> + <edge from-layer="1448" from-port="2" to-layer="1449" to-port="0" /> + <edge from-layer="1449" from-port="2" to-layer="1452" to-port="0" /> + <edge from-layer="1449" from-port="2" to-layer="1451" to-port="0" /> + <edge from-layer="1450" from-port="0" to-layer="1451" to-port="1" /> + <edge from-layer="1451" from-port="2" to-layer="1452" to-port="1" /> + <edge from-layer="1452" from-port="2" to-layer="1460" to-port="0" /> + <edge from-layer="1452" from-port="2" to-layer="1454" to-port="0" /> + <edge from-layer="1453" from-port="0" to-layer="1454" to-port="1" /> + <edge from-layer="1454" from-port="2" to-layer="1456" to-port="0" /> + <edge from-layer="1455" from-port="0" to-layer="1456" to-port="1" /> + <edge from-layer="1456" from-port="2" to-layer="1458" to-port="0" /> + <edge from-layer="1457" from-port="0" to-layer="1458" to-port="1" /> + <edge from-layer="1458" from-port="2" to-layer="1459" to-port="0" /> + <edge from-layer="1459" from-port="1" to-layer="1460" to-port="1" /> + <edge from-layer="1460" from-port="2" to-layer="1462" to-port="0" /> + <edge from-layer="1461" from-port="0" to-layer="1462" to-port="1" /> + <edge from-layer="1462" from-port="2" to-layer="1464" to-port="0" /> + <edge from-layer="1463" from-port="0" to-layer="1464" to-port="1" /> + <edge from-layer="1464" from-port="2" to-layer="1466" to-port="0" /> + <edge from-layer="1464" from-port="2" to-layer="1479" to-port="1" /> + <edge from-layer="1465" from-port="0" to-layer="1466" to-port="1" /> + <edge from-layer="1466" from-port="2" to-layer="1467" to-port="1" /> + <edge from-layer="1467" from-port="2" to-layer="1473" to-port="0" /> + <edge from-layer="1467" from-port="2" to-layer="1469" to-port="0" /> + <edge from-layer="1468" from-port="0" to-layer="1469" to-port="1" /> + <edge from-layer="1469" from-port="2" to-layer="1470" to-port="0" /> + <edge from-layer="1470" from-port="1" to-layer="1472" to-port="0" /> + <edge from-layer="1471" from-port="0" to-layer="1472" to-port="1" /> + <edge from-layer="1472" from-port="2" to-layer="1473" to-port="1" /> + <edge from-layer="1473" from-port="2" to-layer="1475" to-port="0" /> + <edge from-layer="1474" from-port="0" to-layer="1475" to-port="1" /> + <edge from-layer="1475" from-port="2" to-layer="1477" to-port="0" /> + <edge from-layer="1476" from-port="0" to-layer="1477" to-port="1" /> + <edge from-layer="1477" from-port="2" to-layer="1478" to-port="1" /> + <edge from-layer="1478" from-port="2" to-layer="1479" to-port="0" /> + <edge from-layer="1479" from-port="2" to-layer="1482" to-port="0" /> + <edge from-layer="1479" from-port="2" to-layer="1481" to-port="0" /> + <edge from-layer="1480" from-port="0" to-layer="1481" to-port="1" /> + <edge from-layer="1481" from-port="2" to-layer="1482" to-port="1" /> + <edge from-layer="1482" from-port="2" to-layer="1490" to-port="0" /> + <edge from-layer="1482" from-port="2" to-layer="1484" to-port="0" /> + <edge from-layer="1483" from-port="0" to-layer="1484" to-port="1" /> + <edge from-layer="1484" from-port="2" to-layer="1486" to-port="0" /> + <edge from-layer="1485" from-port="0" to-layer="1486" to-port="1" /> + <edge from-layer="1486" from-port="2" to-layer="1488" to-port="0" /> + <edge from-layer="1487" from-port="0" to-layer="1488" to-port="1" /> + <edge from-layer="1488" from-port="2" to-layer="1489" to-port="0" /> + <edge from-layer="1489" from-port="1" to-layer="1490" to-port="1" /> + <edge from-layer="1490" from-port="2" to-layer="1492" to-port="0" /> + <edge from-layer="1491" from-port="0" to-layer="1492" to-port="1" /> + <edge from-layer="1492" from-port="2" to-layer="1494" to-port="0" /> + <edge from-layer="1493" from-port="0" to-layer="1494" to-port="1" /> + <edge from-layer="1494" from-port="2" to-layer="1545" to-port="0" /> + <edge from-layer="1494" from-port="2" to-layer="1586" to-port="1" /> + <edge from-layer="1494" from-port="2" to-layer="1518" to-port="0" /> + <edge from-layer="1494" from-port="2" to-layer="1496" to-port="0" /> + <edge from-layer="1495" from-port="0" to-layer="1496" to-port="1" /> + <edge from-layer="1496" from-port="2" to-layer="1497" to-port="1" /> + <edge from-layer="1497" from-port="2" to-layer="1498" to-port="0" /> + <edge from-layer="1497" from-port="2" to-layer="1513" to-port="0" /> + <edge from-layer="1497" from-port="2" to-layer="1504" to-port="0" /> + <edge from-layer="1498" from-port="1" to-layer="1501" to-port="0" /> + <edge from-layer="1499" from-port="0" to-layer="1501" to-port="1" /> + <edge from-layer="1500" from-port="0" to-layer="1501" to-port="2" /> + <edge from-layer="1501" from-port="3" to-layer="1503" to-port="0" /> + <edge from-layer="1502" from-port="0" to-layer="1503" to-port="1" /> + <edge from-layer="1503" from-port="2" to-layer="1512" to-port="0" /> + <edge from-layer="1504" from-port="1" to-layer="1507" to-port="0" /> + <edge from-layer="1505" from-port="0" to-layer="1507" to-port="1" /> + <edge from-layer="1506" from-port="0" to-layer="1507" to-port="2" /> + <edge from-layer="1507" from-port="3" to-layer="1509" to-port="0" /> + <edge from-layer="1508" from-port="0" to-layer="1509" to-port="1" /> + <edge from-layer="1509" from-port="2" to-layer="1512" to-port="1" /> + <edge from-layer="1510" from-port="0" to-layer="1512" to-port="2" /> + <edge from-layer="1511" from-port="0" to-layer="1512" to-port="3" /> + <edge from-layer="1512" from-port="4" to-layer="1513" to-port="1" /> + <edge from-layer="1513" from-port="2" to-layer="1515" to-port="0" /> + <edge from-layer="1514" from-port="0" to-layer="1515" to-port="1" /> + <edge from-layer="1515" from-port="2" to-layer="1538" to-port="0" /> + <edge from-layer="1516" from-port="0" to-layer="1519" to-port="0" /> + <edge from-layer="1517" from-port="0" to-layer="1518" to-port="1" /> + <edge from-layer="1518" from-port="2" to-layer="1519" to-port="1" /> + <edge from-layer="1519" from-port="2" to-layer="1526" to-port="0" /> + <edge from-layer="1519" from-port="2" to-layer="1520" to-port="0" /> + <edge from-layer="1519" from-port="2" to-layer="1535" to-port="0" /> + <edge from-layer="1520" from-port="1" to-layer="1523" to-port="0" /> + <edge from-layer="1521" from-port="0" to-layer="1523" to-port="1" /> + <edge from-layer="1522" from-port="0" to-layer="1523" to-port="2" /> + <edge from-layer="1523" from-port="3" to-layer="1525" to-port="0" /> + <edge from-layer="1524" from-port="0" to-layer="1525" to-port="1" /> + <edge from-layer="1525" from-port="2" to-layer="1534" to-port="0" /> + <edge from-layer="1526" from-port="1" to-layer="1529" to-port="0" /> + <edge from-layer="1527" from-port="0" to-layer="1529" to-port="1" /> + <edge from-layer="1528" from-port="0" to-layer="1529" to-port="2" /> + <edge from-layer="1529" from-port="3" to-layer="1531" to-port="0" /> + <edge from-layer="1530" from-port="0" to-layer="1531" to-port="1" /> + <edge from-layer="1531" from-port="2" to-layer="1534" to-port="1" /> + <edge from-layer="1532" from-port="0" to-layer="1534" to-port="2" /> + <edge from-layer="1533" from-port="0" to-layer="1534" to-port="3" /> + <edge from-layer="1534" from-port="4" to-layer="1535" to-port="1" /> + <edge from-layer="1535" from-port="2" to-layer="1537" to-port="0" /> + <edge from-layer="1536" from-port="0" to-layer="1537" to-port="1" /> + <edge from-layer="1537" from-port="2" to-layer="1538" to-port="1" /> + <edge from-layer="1538" from-port="2" to-layer="1540" to-port="0" /> + <edge from-layer="1539" from-port="0" to-layer="1540" to-port="1" /> + <edge from-layer="1540" from-port="2" to-layer="1541" to-port="0" /> + <edge from-layer="1541" from-port="2" to-layer="1542" to-port="0" /> + <edge from-layer="1542" from-port="1" to-layer="1565" to-port="0" /> + <edge from-layer="1543" from-port="0" to-layer="1546" to-port="0" /> + <edge from-layer="1544" from-port="0" to-layer="1545" to-port="1" /> + <edge from-layer="1545" from-port="2" to-layer="1546" to-port="1" /> + <edge from-layer="1546" from-port="2" to-layer="1553" to-port="0" /> + <edge from-layer="1546" from-port="2" to-layer="1562" to-port="0" /> + <edge from-layer="1546" from-port="2" to-layer="1547" to-port="0" /> + <edge from-layer="1547" from-port="1" to-layer="1550" to-port="0" /> + <edge from-layer="1548" from-port="0" to-layer="1550" to-port="1" /> + <edge from-layer="1549" from-port="0" to-layer="1550" to-port="2" /> + <edge from-layer="1550" from-port="3" to-layer="1552" to-port="0" /> + <edge from-layer="1551" from-port="0" to-layer="1552" to-port="1" /> + <edge from-layer="1552" from-port="2" to-layer="1561" to-port="0" /> + <edge from-layer="1553" from-port="1" to-layer="1556" to-port="0" /> + <edge from-layer="1554" from-port="0" to-layer="1556" to-port="1" /> + <edge from-layer="1555" from-port="0" to-layer="1556" to-port="2" /> + <edge from-layer="1556" from-port="3" to-layer="1558" to-port="0" /> + <edge from-layer="1557" from-port="0" to-layer="1558" to-port="1" /> + <edge from-layer="1558" from-port="2" to-layer="1561" to-port="1" /> + <edge from-layer="1559" from-port="0" to-layer="1561" to-port="2" /> + <edge from-layer="1560" from-port="0" to-layer="1561" to-port="3" /> + <edge from-layer="1561" from-port="4" to-layer="1562" to-port="1" /> + <edge from-layer="1562" from-port="2" to-layer="1564" to-port="0" /> + <edge from-layer="1563" from-port="0" to-layer="1564" to-port="1" /> + <edge from-layer="1564" from-port="2" to-layer="1565" to-port="1" /> + <edge from-layer="1565" from-port="2" to-layer="1567" to-port="0" /> + <edge from-layer="1566" from-port="0" to-layer="1567" to-port="1" /> + <edge from-layer="1567" from-port="2" to-layer="1568" to-port="0" /> + <edge from-layer="1567" from-port="2" to-layer="1582" to-port="0" /> + <edge from-layer="1567" from-port="2" to-layer="1574" to-port="0" /> + <edge from-layer="1568" from-port="1" to-layer="1571" to-port="0" /> + <edge from-layer="1569" from-port="0" to-layer="1571" to-port="1" /> + <edge from-layer="1570" from-port="0" to-layer="1571" to-port="2" /> + <edge from-layer="1571" from-port="3" to-layer="1573" to-port="0" /> + <edge from-layer="1572" from-port="0" to-layer="1573" to-port="1" /> + <edge from-layer="1573" from-port="2" to-layer="1581" to-port="0" /> + <edge from-layer="1574" from-port="1" to-layer="1577" to-port="0" /> + <edge from-layer="1575" from-port="0" to-layer="1577" to-port="1" /> + <edge from-layer="1576" from-port="0" to-layer="1577" to-port="2" /> + <edge from-layer="1577" from-port="3" to-layer="1579" to-port="0" /> + <edge from-layer="1578" from-port="0" to-layer="1579" to-port="1" /> + <edge from-layer="1579" from-port="2" to-layer="1581" to-port="1" /> + <edge from-layer="1580" from-port="0" to-layer="1581" to-port="2" /> + <edge from-layer="1581" from-port="3" to-layer="1582" to-port="1" /> + <edge from-layer="1582" from-port="2" to-layer="1584" to-port="0" /> + <edge from-layer="1583" from-port="0" to-layer="1584" to-port="1" /> + <edge from-layer="1584" from-port="2" to-layer="1585" to-port="1" /> + <edge from-layer="1585" from-port="2" to-layer="1586" to-port="0" /> + <edge from-layer="1586" from-port="2" to-layer="1589" to-port="0" /> + <edge from-layer="1586" from-port="2" to-layer="1588" to-port="0" /> + <edge from-layer="1587" from-port="0" to-layer="1588" to-port="1" /> + <edge from-layer="1588" from-port="2" to-layer="1589" to-port="1" /> + <edge from-layer="1589" from-port="2" to-layer="1597" to-port="0" /> + <edge from-layer="1589" from-port="2" to-layer="1591" to-port="0" /> + <edge from-layer="1590" from-port="0" to-layer="1591" to-port="1" /> + <edge from-layer="1591" from-port="2" to-layer="1593" to-port="0" /> + <edge from-layer="1592" from-port="0" to-layer="1593" to-port="1" /> + <edge from-layer="1593" from-port="2" to-layer="1595" to-port="0" /> + <edge from-layer="1594" from-port="0" to-layer="1595" to-port="1" /> + <edge from-layer="1595" from-port="2" to-layer="1596" to-port="0" /> + <edge from-layer="1596" from-port="1" to-layer="1597" to-port="1" /> + <edge from-layer="1597" from-port="2" to-layer="1599" to-port="0" /> + <edge from-layer="1598" from-port="0" to-layer="1599" to-port="1" /> + <edge from-layer="1599" from-port="2" to-layer="1601" to-port="0" /> + <edge from-layer="1600" from-port="0" to-layer="1601" to-port="1" /> + <edge from-layer="1601" from-port="2" to-layer="1603" to-port="0" /> + <edge from-layer="1601" from-port="2" to-layer="1616" to-port="1" /> + <edge from-layer="1602" from-port="0" to-layer="1603" to-port="1" /> + <edge from-layer="1603" from-port="2" to-layer="1604" to-port="1" /> + <edge from-layer="1604" from-port="2" to-layer="1610" to-port="0" /> + <edge from-layer="1604" from-port="2" to-layer="1606" to-port="0" /> + <edge from-layer="1605" from-port="0" to-layer="1606" to-port="1" /> + <edge from-layer="1606" from-port="2" to-layer="1607" to-port="0" /> + <edge from-layer="1607" from-port="1" to-layer="1609" to-port="0" /> + <edge from-layer="1608" from-port="0" to-layer="1609" to-port="1" /> + <edge from-layer="1609" from-port="2" to-layer="1610" to-port="1" /> + <edge from-layer="1610" from-port="2" to-layer="1612" to-port="0" /> + <edge from-layer="1611" from-port="0" to-layer="1612" to-port="1" /> + <edge from-layer="1612" from-port="2" to-layer="1614" to-port="0" /> + <edge from-layer="1613" from-port="0" to-layer="1614" to-port="1" /> + <edge from-layer="1614" from-port="2" to-layer="1615" to-port="1" /> + <edge from-layer="1615" from-port="2" to-layer="1616" to-port="0" /> + <edge from-layer="1616" from-port="2" to-layer="1619" to-port="0" /> + <edge from-layer="1616" from-port="2" to-layer="1618" to-port="0" /> + <edge from-layer="1617" from-port="0" to-layer="1618" to-port="1" /> + <edge from-layer="1618" from-port="2" to-layer="1619" to-port="1" /> + <edge from-layer="1619" from-port="2" to-layer="1621" to-port="0" /> + <edge from-layer="1619" from-port="2" to-layer="1627" to-port="0" /> + <edge from-layer="1620" from-port="0" to-layer="1621" to-port="1" /> + <edge from-layer="1621" from-port="2" to-layer="1623" to-port="0" /> + <edge from-layer="1622" from-port="0" to-layer="1623" to-port="1" /> + <edge from-layer="1623" from-port="2" to-layer="1625" to-port="0" /> + <edge from-layer="1624" from-port="0" to-layer="1625" to-port="1" /> + <edge from-layer="1625" from-port="2" to-layer="1626" to-port="0" /> + <edge from-layer="1626" from-port="1" to-layer="1627" to-port="1" /> + <edge from-layer="1627" from-port="2" to-layer="1629" to-port="0" /> + <edge from-layer="1628" from-port="0" to-layer="1629" to-port="1" /> + <edge from-layer="1629" from-port="2" to-layer="1631" to-port="0" /> + <edge from-layer="1630" from-port="0" to-layer="1631" to-port="1" /> + <edge from-layer="1631" from-port="2" to-layer="1723" to-port="1" /> + <edge from-layer="1631" from-port="2" to-layer="1682" to-port="0" /> + <edge from-layer="1631" from-port="2" to-layer="1655" to-port="0" /> + <edge from-layer="1631" from-port="2" to-layer="1633" to-port="0" /> + <edge from-layer="1632" from-port="0" to-layer="1633" to-port="1" /> + <edge from-layer="1633" from-port="2" to-layer="1634" to-port="1" /> + <edge from-layer="1634" from-port="2" to-layer="1641" to-port="0" /> + <edge from-layer="1634" from-port="2" to-layer="1635" to-port="0" /> + <edge from-layer="1634" from-port="2" to-layer="1650" to-port="0" /> + <edge from-layer="1635" from-port="1" to-layer="1638" to-port="0" /> + <edge from-layer="1636" from-port="0" to-layer="1638" to-port="1" /> + <edge from-layer="1637" from-port="0" to-layer="1638" to-port="2" /> + <edge from-layer="1638" from-port="3" to-layer="1640" to-port="0" /> + <edge from-layer="1639" from-port="0" to-layer="1640" to-port="1" /> + <edge from-layer="1640" from-port="2" to-layer="1649" to-port="0" /> + <edge from-layer="1641" from-port="1" to-layer="1644" to-port="0" /> + <edge from-layer="1642" from-port="0" to-layer="1644" to-port="1" /> + <edge from-layer="1643" from-port="0" to-layer="1644" to-port="2" /> + <edge from-layer="1644" from-port="3" to-layer="1646" to-port="0" /> + <edge from-layer="1645" from-port="0" to-layer="1646" to-port="1" /> + <edge from-layer="1646" from-port="2" to-layer="1649" to-port="1" /> + <edge from-layer="1647" from-port="0" to-layer="1649" to-port="2" /> + <edge from-layer="1648" from-port="0" to-layer="1649" to-port="3" /> + <edge from-layer="1649" from-port="4" to-layer="1650" to-port="1" /> + <edge from-layer="1650" from-port="2" to-layer="1652" to-port="0" /> + <edge from-layer="1651" from-port="0" to-layer="1652" to-port="1" /> + <edge from-layer="1652" from-port="2" to-layer="1675" to-port="0" /> + <edge from-layer="1653" from-port="0" to-layer="1656" to-port="0" /> + <edge from-layer="1654" from-port="0" to-layer="1655" to-port="1" /> + <edge from-layer="1655" from-port="2" to-layer="1656" to-port="1" /> + <edge from-layer="1656" from-port="2" to-layer="1657" to-port="0" /> + <edge from-layer="1656" from-port="2" to-layer="1672" to-port="0" /> + <edge from-layer="1656" from-port="2" to-layer="1663" to-port="0" /> + <edge from-layer="1657" from-port="1" to-layer="1660" to-port="0" /> + <edge from-layer="1658" from-port="0" to-layer="1660" to-port="1" /> + <edge from-layer="1659" from-port="0" to-layer="1660" to-port="2" /> + <edge from-layer="1660" from-port="3" to-layer="1662" to-port="0" /> + <edge from-layer="1661" from-port="0" to-layer="1662" to-port="1" /> + <edge from-layer="1662" from-port="2" to-layer="1671" to-port="0" /> + <edge from-layer="1663" from-port="1" to-layer="1666" to-port="0" /> + <edge from-layer="1664" from-port="0" to-layer="1666" to-port="1" /> + <edge from-layer="1665" from-port="0" to-layer="1666" to-port="2" /> + <edge from-layer="1666" from-port="3" to-layer="1668" to-port="0" /> + <edge from-layer="1667" from-port="0" to-layer="1668" to-port="1" /> + <edge from-layer="1668" from-port="2" to-layer="1671" to-port="1" /> + <edge from-layer="1669" from-port="0" to-layer="1671" to-port="2" /> + <edge from-layer="1670" from-port="0" to-layer="1671" to-port="3" /> + <edge from-layer="1671" from-port="4" to-layer="1672" to-port="1" /> + <edge from-layer="1672" from-port="2" to-layer="1674" to-port="0" /> + <edge from-layer="1673" from-port="0" to-layer="1674" to-port="1" /> + <edge from-layer="1674" from-port="2" to-layer="1675" to-port="1" /> + <edge from-layer="1675" from-port="2" to-layer="1677" to-port="0" /> + <edge from-layer="1676" from-port="0" to-layer="1677" to-port="1" /> + <edge from-layer="1677" from-port="2" to-layer="1678" to-port="0" /> + <edge from-layer="1678" from-port="2" to-layer="1679" to-port="0" /> + <edge from-layer="1679" from-port="1" to-layer="1702" to-port="0" /> + <edge from-layer="1680" from-port="0" to-layer="1683" to-port="0" /> + <edge from-layer="1681" from-port="0" to-layer="1682" to-port="1" /> + <edge from-layer="1682" from-port="2" to-layer="1683" to-port="1" /> + <edge from-layer="1683" from-port="2" to-layer="1690" to-port="0" /> + <edge from-layer="1683" from-port="2" to-layer="1684" to-port="0" /> + <edge from-layer="1683" from-port="2" to-layer="1699" to-port="0" /> + <edge from-layer="1684" from-port="1" to-layer="1687" to-port="0" /> + <edge from-layer="1685" from-port="0" to-layer="1687" to-port="1" /> + <edge from-layer="1686" from-port="0" to-layer="1687" to-port="2" /> + <edge from-layer="1687" from-port="3" to-layer="1689" to-port="0" /> + <edge from-layer="1688" from-port="0" to-layer="1689" to-port="1" /> + <edge from-layer="1689" from-port="2" to-layer="1698" to-port="0" /> + <edge from-layer="1690" from-port="1" to-layer="1693" to-port="0" /> + <edge from-layer="1691" from-port="0" to-layer="1693" to-port="1" /> + <edge from-layer="1692" from-port="0" to-layer="1693" to-port="2" /> + <edge from-layer="1693" from-port="3" to-layer="1695" to-port="0" /> + <edge from-layer="1694" from-port="0" to-layer="1695" to-port="1" /> + <edge from-layer="1695" from-port="2" to-layer="1698" to-port="1" /> + <edge from-layer="1696" from-port="0" to-layer="1698" to-port="2" /> + <edge from-layer="1697" from-port="0" to-layer="1698" to-port="3" /> + <edge from-layer="1698" from-port="4" to-layer="1699" to-port="1" /> + <edge from-layer="1699" from-port="2" to-layer="1701" to-port="0" /> + <edge from-layer="1700" from-port="0" to-layer="1701" to-port="1" /> + <edge from-layer="1701" from-port="2" to-layer="1702" to-port="1" /> + <edge from-layer="1702" from-port="2" to-layer="1704" to-port="0" /> + <edge from-layer="1703" from-port="0" to-layer="1704" to-port="1" /> + <edge from-layer="1704" from-port="2" to-layer="1719" to-port="0" /> + <edge from-layer="1704" from-port="2" to-layer="1711" to-port="0" /> + <edge from-layer="1704" from-port="2" to-layer="1705" to-port="0" /> + <edge from-layer="1705" from-port="1" to-layer="1708" to-port="0" /> + <edge from-layer="1706" from-port="0" to-layer="1708" to-port="1" /> + <edge from-layer="1707" from-port="0" to-layer="1708" to-port="2" /> + <edge from-layer="1708" from-port="3" to-layer="1710" to-port="0" /> + <edge from-layer="1709" from-port="0" to-layer="1710" to-port="1" /> + <edge from-layer="1710" from-port="2" to-layer="1718" to-port="0" /> + <edge from-layer="1711" from-port="1" to-layer="1714" to-port="0" /> + <edge from-layer="1712" from-port="0" to-layer="1714" to-port="1" /> + <edge from-layer="1713" from-port="0" to-layer="1714" to-port="2" /> + <edge from-layer="1714" from-port="3" to-layer="1716" to-port="0" /> + <edge from-layer="1715" from-port="0" to-layer="1716" to-port="1" /> + <edge from-layer="1716" from-port="2" to-layer="1718" to-port="1" /> + <edge from-layer="1717" from-port="0" to-layer="1718" to-port="2" /> + <edge from-layer="1718" from-port="3" to-layer="1719" to-port="1" /> + <edge from-layer="1719" from-port="2" to-layer="1721" to-port="0" /> + <edge from-layer="1720" from-port="0" to-layer="1721" to-port="1" /> + <edge from-layer="1721" from-port="2" to-layer="1722" to-port="1" /> + <edge from-layer="1722" from-port="2" to-layer="1723" to-port="0" /> + <edge from-layer="1723" from-port="2" to-layer="1726" to-port="0" /> + <edge from-layer="1723" from-port="2" to-layer="1725" to-port="0" /> + <edge from-layer="1724" from-port="0" to-layer="1725" to-port="1" /> + <edge from-layer="1725" from-port="2" to-layer="1726" to-port="1" /> + <edge from-layer="1726" from-port="2" to-layer="1734" to-port="0" /> + <edge from-layer="1726" from-port="2" to-layer="1728" to-port="0" /> + <edge from-layer="1727" from-port="0" to-layer="1728" to-port="1" /> + <edge from-layer="1728" from-port="2" to-layer="1730" to-port="0" /> + <edge from-layer="1729" from-port="0" to-layer="1730" to-port="1" /> + <edge from-layer="1730" from-port="2" to-layer="1732" to-port="0" /> + <edge from-layer="1731" from-port="0" to-layer="1732" to-port="1" /> + <edge from-layer="1732" from-port="2" to-layer="1733" to-port="0" /> + <edge from-layer="1733" from-port="1" to-layer="1734" to-port="1" /> + <edge from-layer="1734" from-port="2" to-layer="1736" to-port="0" /> + <edge from-layer="1735" from-port="0" to-layer="1736" to-port="1" /> + <edge from-layer="1736" from-port="2" to-layer="1738" to-port="0" /> + <edge from-layer="1737" from-port="0" to-layer="1738" to-port="1" /> + <edge from-layer="1738" from-port="2" to-layer="1740" to-port="0" /> + <edge from-layer="1738" from-port="2" to-layer="1753" to-port="1" /> + <edge from-layer="1739" from-port="0" to-layer="1740" to-port="1" /> + <edge from-layer="1740" from-port="2" to-layer="1741" to-port="1" /> + <edge from-layer="1741" from-port="2" to-layer="1747" to-port="0" /> + <edge from-layer="1741" from-port="2" to-layer="1743" to-port="0" /> + <edge from-layer="1742" from-port="0" to-layer="1743" to-port="1" /> + <edge from-layer="1743" from-port="2" to-layer="1744" to-port="0" /> + <edge from-layer="1744" from-port="1" to-layer="1746" to-port="0" /> + <edge from-layer="1745" from-port="0" to-layer="1746" to-port="1" /> + <edge from-layer="1746" from-port="2" to-layer="1747" to-port="1" /> + <edge from-layer="1747" from-port="2" to-layer="1749" to-port="0" /> + <edge from-layer="1748" from-port="0" to-layer="1749" to-port="1" /> + <edge from-layer="1749" from-port="2" to-layer="1751" to-port="0" /> + <edge from-layer="1750" from-port="0" to-layer="1751" to-port="1" /> + <edge from-layer="1751" from-port="2" to-layer="1752" to-port="1" /> + <edge from-layer="1752" from-port="2" to-layer="1753" to-port="0" /> + <edge from-layer="1753" from-port="2" to-layer="1755" to-port="0" /> + <edge from-layer="1753" from-port="2" to-layer="1756" to-port="0" /> + <edge from-layer="1754" from-port="0" to-layer="1755" to-port="1" /> + <edge from-layer="1755" from-port="2" to-layer="1756" to-port="1" /> + <edge from-layer="1756" from-port="2" to-layer="1758" to-port="0" /> + <edge from-layer="1756" from-port="2" to-layer="1764" to-port="0" /> + <edge from-layer="1757" from-port="0" to-layer="1758" to-port="1" /> + <edge from-layer="1758" from-port="2" to-layer="1760" to-port="0" /> + <edge from-layer="1759" from-port="0" to-layer="1760" to-port="1" /> + <edge from-layer="1760" from-port="2" to-layer="1762" to-port="0" /> + <edge from-layer="1761" from-port="0" to-layer="1762" to-port="1" /> + <edge from-layer="1762" from-port="2" to-layer="1763" to-port="0" /> + <edge from-layer="1763" from-port="1" to-layer="1764" to-port="1" /> + <edge from-layer="1764" from-port="2" to-layer="1766" to-port="0" /> + <edge from-layer="1765" from-port="0" to-layer="1766" to-port="1" /> + <edge from-layer="1766" from-port="2" to-layer="1768" to-port="0" /> + <edge from-layer="1767" from-port="0" to-layer="1768" to-port="1" /> + <edge from-layer="1768" from-port="2" to-layer="1771" to-port="0" /> + <edge from-layer="1769" from-port="0" to-layer="1771" to-port="1" /> + <edge from-layer="1770" from-port="0" to-layer="1771" to-port="2" /> + <edge from-layer="1771" from-port="3" to-layer="1773" to-port="0" /> + <edge from-layer="1772" from-port="0" to-layer="1773" to-port="1" /> + <edge from-layer="1773" from-port="2" to-layer="1775" to-port="0" /> + <edge from-layer="1774" from-port="0" to-layer="1775" to-port="1" /> + <edge from-layer="1775" from-port="2" to-layer="1779" to-port="0" /> + <edge from-layer="1776" from-port="0" to-layer="1778" to-port="0" /> + <edge from-layer="1777" from-port="0" to-layer="1778" to-port="1" /> + <edge from-layer="1778" from-port="2" to-layer="1779" to-port="1" /> + <edge from-layer="1779" from-port="2" to-layer="1780" to-port="0" /> + <edge from-layer="1780" from-port="1" to-layer="1782" to-port="0" /> + <edge from-layer="1781" from-port="0" to-layer="1782" to-port="1" /> + <edge from-layer="1782" from-port="2" to-layer="1784" to-port="0" /> + <edge from-layer="1783" from-port="0" to-layer="1784" to-port="1" /> + <edge from-layer="1784" from-port="2" to-layer="1788" to-port="0" /> + <edge from-layer="1785" from-port="0" to-layer="1787" to-port="0" /> + <edge from-layer="1786" from-port="0" to-layer="1787" to-port="1" /> + <edge from-layer="1787" from-port="2" to-layer="1788" to-port="1" /> + <edge from-layer="1788" from-port="2" to-layer="1789" to-port="0" /> + </edges> + <rt_info /> +</net>